2024.01.17.576140v1.full-html.html

Gut Analysis Toolbox

Gut Analysis Toolbox: Automating

quantitative analysis of enteric neurons

Luke Sorensen

, Adam Humenick

, Sabrina S.B. Poon

, Myat Noe Han

, Narges Sadat

Mahdavian

, Ryan Hamnett

4,5

, Estibaliz Gómez-de-Mariscal

, Peter H. Neckel

, Ayame

Saito

, Keith Mutunduwe

, Christie Glennan

, Robert Haase

, Rachel M. McQuade

9,10,11

Jaime P.P. Foong

, Simon J.H. Brookes

, Julia A. Kaltschmidt

4,5

, Arrate Muñoz-Barrutia

12,13

Sebastian K. King

14,15,16

, Nicholas A. Veldhuis

, Simona E. Carbone

, Daniel P. Poole

Pradeep Rajasekhar

17,18

Drug Discovery Biology, Monash Institute of Pharmaceutical Sciences, Monash University,

Parkville, Victoria 3052, Australia

Flinders Health and Medical Research Institute, College of Medicine and Public Health,

Flinders University, Bedford Park, SA, 5042, Australia

Department of Anatomy and Physiology, University of Melbourne, Parkville, Victoria

3052, Australia

Wu Tsai Neurosciences Institute, Stanford University, Stanford, CA 94305, USA

Department of Neurosurgery, Stanford University School of Medicine, Stanford, CA 94305,

USA

Instituto Gulbenkian de Ciência, Oeiras 2780-156, Portugal

Institute of Clinical Anatomy and Cell Analysis, University of Tübingen, Tübingen,

Germany

Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Dresden/Leipzig,

Universität Leipzig, Germany

Gut Barrier and Disease Laboratory, Department of Anatomy and Physiology, The

University of Melbourne, Melbourne, Victoria, Australia.

Department of Medicine, Western Health, The University of Melbourne, Melbourne,

Victoria, Australia.

Australian Institute for Musculoskeletal Science (AIMSS), Melbourne University,

Melbourne, Victoria, Australia.

Bioengineering Department, Universidad Carlos III de Madrid, ES28911-Leganés, Spain

Bioengineering Division, Instituto de Investigación Sanitaria Gregorio Marañon, ES28007-

Madrid, Spain

Department of Paediatric Surgery, The Royal Children's Hospital, Parkville, VIC 3052,

Australia

Surgical Research, Murdoch Children's Research Institute, Parkville, VIC 3052, Australia

Department of Paediatrics, The University of Melbourne, Parkville, VIC 3010, Australia

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

The Walter and Eliza Hall Institute of Medical Research, Parkville, Victoria, 3052,

Australia

Department of Medical Biology, The University of Melbourne, Parkville, Victoria, 3052,

Australia

Abstract:

The enteric nervous system (ENS) plays an important role in coordinating gut function. The

ENS consists of an extensive network of neurons and glial cells within the wall of the

gastrointestinal tract. Alterations in neuronal distribution, function, and type are strongly

associated with enteric neuropathies and gastrointestinal (GI) dysfunction and can serve as

biomarkers for disease. However, current methods for assessing neuronal counts and

distribution suffer from undersampling. This is partly due to challenges associated with

imaging and analyzing large tissue areas, and operator bias due to manual analysis. Here, we

present the Gut Analysis Toolbox (GAT), an image analysis tool designed for

characterization of enteric neurons and their neurochemical coding using 2D images of GI

wholemount preparations. GAT is developed for the Fiji distribution of ImageJ. It has a user-

friendly interface and offers rapid and accurate cell segmentation. Custom deep learning (DL)

based cell segmentation models were developed using StarDist. GAT also includes a

ganglion segmentation model which was developed using deepImageJ. In addition, GAT

allows importing of segmentation generated by other software. DL models have been trained

using ZeroCostDL4Mic on diverse datasets sourced from different laboratories. This captures

the variability associated with differences in animal species, image acquisition parameters,

and sample preparation across research groups. We demonstrate the robustness of the cell

segmentation DL models by comparing them against the state-of-the-art cell segmentation

software, Cellpose. To quantify neuronal distribution GAT applies proximal neighbor-based

spatial analysis. We demonstrate how the proximal neighbor analysis can reveal differences

in cellular distribution across gut regions using a published dataset. In summary, GAT

provides an easy-to-use toolbox to streamline routine image analysis tasks in ENS research.

GAT enhances throughput allowing unbiased analysis of larger tissue areas, multiple

neuronal markers and numerous samples rapidly.

Introduction:

The enteric nervous system (ENS) is a network of neurons and glial cells located within the

wall of the gastrointestinal (GI) tract. The ENS extends along the esophagus to the rectum

and is estimated to be comprised of ~168 million neurons, which is comparable to the

numbers in the spinal cord in humans, mice, and guinea pigs (Michel et al. 2022a). It is

critical for the regulation of secretion, absorption, and immune function, and for coordination

of gut motility (Furness 2012). The absence or loss of enteric neurons results in GI

dysfunction as evidenced in enteric neuropathies such as Hirschsprung disease, achalasia, and

Chagas disease (Burns et al. 2016; Heuckeroth 2018; Schäppi et al. 2013; Vaezi et al. 2016).

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Alterations to specific enteric neuron populations that express distinct combinations of

neuropeptides, enzymes or neurochemicals are also evident in other diseases that impact gut

function. These include inflammatory bowel disease (Brierley & Linden 2014), diabetes

(Chandrasekharan et al. 2011; Ingrid et al. 2013), and Alzheimer’s disease (Niesler et al.

2021; Semar et al. 2013; Van Ginneken et al. 2011). Researchers use enteric neuronal counts

as a key metric to describe any neurochemical changes in these diseases. Typically, this is

achieved by manually counting cells in small intestinal segments, or more recently, via semi-

automated methods (Cairns et al. 2021; Kapur 2013; Kobayashi et al. 2021; Nestor-Kalinoski

et al. 2022; Schäppi et al. 2013). These cell counts from localized areas within a specimen are

then used to make inferences about broader changes to the bowel region being studied and

any changes associated with disease. However, the estimated number of cells counted can be

affected by:

•

Tissue preparation examined: Cell density estimates using tissue sections can be

prone to sampling and operator bias compared to the use of wholemounts (Kapur

2013; Swaminathan & Kapur 2010).

•

Age of the animal (Gamage et al. 2013).

•

Tissue region examined: Regional differences in the distribution of ENS circuitry

(Hamnett, Dershowitz, Sampathkumar, et al. 2022; Nestor-Kalinoski et al. 2022).

•

Sampling: The number of tissue specimens and locations sampled (Nestor-Kalinoski

et al. 2022).

•

Operator bias: Bias during the tissue preparation, sampling, or manual counting steps

(Kapur 2013; Schäppi et al. 2013).

The major limitation of current approaches for neuronal quantification in large tissue

specimens is the use of manual cell counting. This process is slow, labor intensive, and prone

to inter-observer variability. In our opinion, the continued use of manual counting processes

100

is largely due to the lack of easy-to-use, ENS-specific image analysis software.

101

102

The need for image analysis software in neurogastroenterology is evidenced by the increasing

103

number of image analysis workflows and tools becoming available in recent years, such as

104

COUNTEN and the use of machine learning approaches in Fiji (Cairns et al. 2021;

105

Kobayashi et al. 2021). However, to use these workflows computational expertise is required,

106

and the software parameters need to be optimized for new data sets. COUNTEN is highly

107

dependent on images of robust and homogeneous staining, which is not always achievable for

108

all intestinal preparations (Kobayashi et al. 2021). Multiple variables can influence image

109

quality, such as low signal-to-noise ratio, variations in sample preparation, quality of

110

dissection of layers which can affect antibody penetration and visualization of the ENS, and

111

the specific antibodies, markers, and fluorophores used during staining. Furthermore, image

112

acquisition specifications, including the bit-depth and dynamic range, can significantly affect

113

downstream image analysis. All these factors can pose challenges to the widespread adoption

114

of such software, necessitating the development of customized analytical workflows for each

115

use case.

116

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

117

We developed the Gut Analysis Toolbox (GAT) for the Fiji distribution of ImageJ

118

(Schindelin et al. 2012). GAT can be used to analyze and quantify cells within the ENS. GAT

119

uses deep learning-based (DL) cell segmentation models developed with StarDist (Schmidt et

120

al. 2018) for segmenting enteric neurons and neuronal subtypes. A pre-trained TensorFlow

121

model was used to segment ganglia, which is accessible in Fiji using deepImageJ (Gómez-de-

122

Mariscal et al. 2021). These models are integrated into GAT for rapid and reproducible

123

quantification of key metrics such as total neuronal counts, neurochemical marker

124

distribution, and cell number per ganglion (Fig. 1). The DL models were trained on manually

125

annotated data from mouse, rat, and human colon wholemounts to ensure they effectively

126

segmented a wide variety of images (Chen et al. 2023; DiCello et al. 2020; Graham et al.

127

2020b; Nestor-Kalinoski et al. 2022). Images were acquired using confocal and widefield

128

microscopes from different research groups. Proximal neighbor analysis was used to

129

characterize neuronal distribution using CLIJ (Haase, Jain, et al. 2020; Haase, Royer, et al.

130

2020). Comprehensive installation and usage instructions, along with sample data and tutorial

131

videos, can be found in the documentation available at

https://gut-analysis-

132

toolbox.gitbook.io/docs/

. The GAT workflow is also usable from within the QuPath software

133

to enable analysis of large 2D images (Bankhead et al. 2017). The enhanced throughput of

134

GAT facilitates sampling and analysis of larger tissue areas, thus minimizing potential

135

sampling errors and biases.

136

137

138

Figure 1: Overview of GAT workflow: GAT can segment neurons, neurons expressing

139

neurochemical markers, and ganglia in fluorescently labeled 2D images using pre-trained DL models.

140

GAT allows manual verification of the segmentation, followed by automated quantification of cell

141

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

counts. The cellular distribution can be subsequently quantified via proximal neighbor analysis. Large

142

2D images can be analyzed in QuPath using the GAT DL models. Extensive documentation and

143

videos on how to use GAT are available at:

https://gut-analysis-toolbox.gitbook.io/docs/

. Myenteric

144

wholemount images from DiCello et al. (2020) and Howard (2021) (Figure created using

145

Biorender.com).

146

147

Results

148

149

To develop effective deep learning models, it is essential to have diverse datasets that

150

encompass the inherent variability in images from various sources. To meet this need, we

151

collected ENS images from four different research laboratories (Chen et al. 2023; DiCello et

152

al. 2020; McQuade et al. 2021; Poon et al. 2022) and two publicly available datasets (Graham

153

et al. 2020a; Howard 2021). Images were captured using a variety of microscopes, and the

154

labeled tissues originated from different animal species, including mice, humans, and rats, as

155

detailed in Table S1. The details for images with pan-neuronal marker Hu are summarized in

156

Table S1. For the enteric neuron subtype model (Table S2), nine neurochemical markers

157

were used, with 42% of the images representing nNOS. The ganglia model was trained using

158

various neuronal markers in combination with Hu, as listed in Table S3. The effects of

159

sample preparation and inadequate sampling on enteric neuron counts were also tested, see

160

the supplementary information section (Figures S1, S2 and Supp. Text).

161

162

163

Segmentation models

164

165

StarDist (Schmidt et al. 2018) was used to train DL models for segmenting all enteric

166

neurons. A U-Net architecture was utilized for the ganglia model (Ronneberger, Fischer &

167

Brox 2015). The training and evaluation of the models were conducted using

168

ZeroCostDL4Mic (von Chamier et al. 2021). Further details on the curation of training data

169

and software versions used can be found in the methods and supplementary methods section.

170

The ZeroCostDL4Mic Google Colab notebooks used for training and quality control are

171

attached as supplementary files, and available online (https://zenodo.org/records/6096664).

172

173

174

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

175

176

Figure 2. GAT segments neurons and ganglia with high accuracy: (A) Segmentation of Hu labeled

177

enteric neurons by the GAT StarDist model (Table S1) and Cellpose (cyto2 model) had a comparable

178

F1-score. However, the GAT StarDist model for other neurochemical markers (Table S2) had a better

179

F1 score at various Intersection Over Union (IoU) thresholds. Representative images of different IoU

180

thresholds are depicted as cell outlines below the graph (Scale bar = 5 µ m). (B) The StarDist model

181

could be used to detect bright versus dim cells and overlapping cells. Green outlines highlight GAT-

182

detected cells. The orange outline shows a false positive cell, and the orange arrowhead indicates a

183

missed cell (scale bar = 20 µm; pixel size of 0.5µm/pixel was used for segmentation). (C) When used

184

to count enteric neurons, the lowest percentage of error was achieved using GAT StarDist model

185

compared to Cellpose models and COUNTEN (20 images, 3830 cells). (D) The StarDist enteric

186

neuron subtype model produced a lower percentage error compared to Cellpose when used to count

187

different enteric neurons based on the neurochemical marker expressed (15 images, 359 cells). (E)

188

The total time taken to segment 1309 neurons was significantly faster using GAT compared to manual

189

segmentation. (F, G) A 2D U-Net model was trained to segment ganglia based on the presence of

190

various markers (Table S3) that label the neuron or glial fibers. The representative image

191

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

demonstrates prediction and segmentation using Hu and GFAP. (H) The ganglia segmentation model

192

has a mean IoU of 0.72 ± 0.17 when using a fixed threshold of 0.8, and 0.78 ± 0.09 with a deepImageJ

193

optimized threshold for each image. COUNTEN had a lower IoU of 0.645 ± 0.16 (n = 20 images,

194

scale bar = 30 µ m).

195

196

The performance of the ‘enteric neuron model’ was evaluated on a test dataset, and compared

197

to a widely used cell segmentation software, Cellpose (v0.7) (Stringer et al. 2021). Fig. 2A

198

presents the evaluation of object detection accuracy (F1 score) and shape alignment accuracy

199

(intersection over union, IoU) (Caicedo et al. 2019). A higher F1 score at a higher IoU

200

indicates better segmentation performance (Fig. 2A). The Hu segmentation results showed

201

comparable performance between the StarDist and Cellpose (cyto2) models. However, when

202

examining the neuronal subtype models, a rightward shift in F1 scores of the StarDist model

203

indicates a modest improvement when compared to the Cellpose cyto2 model (Fig. 2A). As

204

StarDist approximates the shape of a cell using star-convex polygons, the predicted objects

205

have smoother outlines relative to the original cell shape (Fig. S5).

206

207

The primary goal of GAT is to estimate cell counts and this metric is not necessarily captured

208

by the F1 score which evaluates segmentation quality. The StarDist models approximate cell

209

shape which leads to smoother outlines, meaning the final segmentation result will slightly

210

differ from the ground truth leading to reduced F1 scores (Fig. S5). The ‘percentage cell

211

count error’, which is the ‘(predicted number - ground truth number) /ground truth number x

212

100’, is a more direct measure of the accuracy of these DL models for evaluating cell count,

213

compared to the F1 score. Importantly, this approach allows for an objective comparison with

214

COUNTEN using the default settings recommended by the authors (Kobayashi et al. 2021).

215

In this context, a lower percentage error indicates enhanced performance. The percentage cell

216

count error was estimated using the same test datasets as those used for the F1 score. The

217

StarDist neuron model had significantly better accuracy with only 6.09 ± 4.8 % error in cell

218

counts, compared to 14.6 ± 12.1 % for Cellpose (20 images, 3830 neurons, Mean ± SD, one-

219

way ANOVA). The percentage error for COUNTEN (Fig. 2C) was also higher than GAT at

220

13.54 ± 10 %. When testing the accuracy of segmenting enteric neurons expressing the

221

markers calbindin (Calb), calretinin (Calret), neuronalNOS (nNOS) and neurofilament M

222

(NFM), the StarDist model demonstrated a similar cell count percentage error (13.8 ± 13.1

223

%) compared to Cellpose (29.5 ± 29.3 %) (Fig. 2D). Notably, being a generalist cell

224

segmentation algorithm, Cellpose still had high IoU scores on the enteric neuron

225

segmentation task, with higher cell count accuracies for 2D confocal images (Fig. 2C). The

226

Cellpose predictions had higher cell splitting and merging errors (Wolny et al. 2020)

227

compared to StarDist (Fig. S6). Training the Cellpose cyto2 model on these enteric neuron

228

datasets will most likely produce a better performing Cellpose enteric neuronal model

229

(Pachitariu & Stringer 2022). However, this was not explored due to the lack of a standalone

230

Fiji plugin for Cellpose.

231

232

Semi-automation using GAT increased analysis throughput. Manual segmentation of 1309

233

enteric neurons took 183 minutes in total, across 3 researchers, whereas segmentation using

234

GAT and the enteric neuron DL model took only 10.4 minutes by a single person (Fig. 2E).

235

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

236

To enable segmentation of ganglia using GAT, a 2D U-Net model was trained using

237

ZeroCostDL4Mic(von Chamier et al. 2021) on images of Hu labelling in combination with a

238

second marker for neuronal or glial fibers (Table S3). The ganglia model was subsequently

239

exported in a format compatible for use within the Fiji deepImageJ plugin (Gómez-de-

240

Mariscal et al. 2021). A limitation is that the postprocessing threshold value applied to the

241

probability map output from deepImageJ can impact the accuracy of the ganglia outline. To

242

evaluate this an arbitrary threshold of 0.8 was compared with the default deepImageJ

243

optimized threshold. The GAT ganglia model performed better compared to COUNTEN

244

when measuring the IoU, particularly when using the optimized threshold (Fig. 2F) (GAT:

245

IoU of 0.78 ± 0.09 for optimized threshold; 0.72 ± 0.17 for a fixed threshold of 0.8, vs

246

COUNTEN: 0.64 ± 0.16, Mean ± SD). Not all laboratories use markers to label the

247

ganglionic border. In this instance, the neuronal outline can be expanded by a user-specified

248

distance to approximate ganglionic area.

249

250

251

252

253

254

255

256

257

258

259

260

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

261

262

263

Fig. 3: Spatial analysis in GAT can adjust for non-uniform tissue stretching and can objectively

264

describe region-specific differences in neuronal distribution: (A) Both panels show the same field

265

of view (FOV) of a myenteric plexus wholemount preparation of the Wnt1-cre:Rosa26-tdTomato

266

mouse colon. The first panel shows the specimen that was pinned in a petri dish with ‘no stretch’ and

267

the second panel is in the presence of ‘stretch’. Stretching the specimen leads to altered calculations of

268

neuronal density (2696 neurons/mm

vs 1493 neurons/mm

) (B). (C) ‘Proximal neighbor’ (PN)

269

analysis shows similar distribution results for non-stretched versus stretched preparations. (D)

270

Proximal neighbor distribution for the proximal colon (PC) vs distal colon (DC) without

271

normalization to the total cell count shows that PC has larger neuronal clusters with with significantly

272

larger neuronal clusters at neighbor counts of 5 and 6. (E) Normalizing for total neuron count (Hu+

273

neurons) revealed that the DC had significantly smaller neuronal clusters at neighbor count of 3,

274

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

whereas PC had significantly larger neuronal clusters at neighbor counts of 5 and 6. The difference in

275

number of neighbors across regions which can be visualized in a neighbor count map (F). (G, H)

276

Calb+ positive neurons were significantly greater in proportion in the PC compared to DC, whereas

277

there was no significant difference in the distribution of Calret+ neurons (I) or Calret+/ Calb+ double-

278

positive neurons (J). This was reflected in the normalized PN distribution plots, where a significantly

279

higher proportion of neurons had one Calb+ neighbor in the PC relative to the DC. No regional

280

differences in Calret+ neurons (K) or preferential association between Calret+ and Calb+ neurons

281

were detected (L, M, N) (Scale bar = 100 µ m).

282

283

Proximal neighbor analysis in GAT can detect regional differences in cellular

284

distribution

285

286

An accurate estimation of neuronal densities in gut wholemounts can be impacted by the

287

degree of stretch applied during tissue processing. Nonetheless, the stretch applied does not

288

alter the cellular architecture and spatial relationships between cells. Any changes in cell

289

density results in a proportional change in the number of neighbors around each cell. As the

290

spatial relationship is unaffected by tissue stretch, the number of proximal neighbors (PNs) is

291

a robust measure for characterizing cellular distribution. To apply PN measurements in GAT,

292

a threshold distance value was determined to define the distance within which cells were

293

considered neighbors. The edge-to-edge distances between the segmented neurons in the

294

ganglia were measured in images of the myenteric wholemounts of the mouse colon (Table

295

S1, & S3). To define a threshold distance value for considering a cell a PN, the average PN

296

distance between neurons (edge to edge distance) in the ganglia was calculated using the

297

‘Local Thickness’ plugin (Dougherty & Kunzelmann 2007) in Fiji (Fig. S7). For the

298

myenteric wholemounts of the mouse colon, this was determined to be 6.32 ± 5.17 µm

299

(n=3643 cells, 130 ganglia, Mean ± S.D., N=11), which was rounded up to 6.5 µm for use

300

within GAT (Fig. S7 C & D). This value is customizable in GAT to account for differences

301

in tissue preparation or varying cell sizes across species. The neighbor count map allowed the

302

examination of any changes associated with different distance values.

303

To evaluate the robustness of spatial analysis with uneven tissue stretch, an image of the

304

same region of a myenteric wholemount of the Wnt1-cre:Rosa26-tdTomato mouse colon was

305

acquired under poorly stretched and stretched conditions (Fig. 3A). Stretch led to a ~55.3 %

306

increase in the tissue area examined (152318 µm

vs. 245652 µm

for unstretched), total

307

ganglionic area (26573 µm

vs. 44844 µm

) and cell densities (2696 neurons/mm

vs. 1493

308

neurons/mm

, Fig. 3A). Using the PN analysis, no significant difference was found between

309

the average PNs around each neuron for both stretched and unstretched tissue 2.76 ± 1.25 vs

310

2.98± 1.21 (Mean ± S.D., p=0.28 Students t-test, N=1). Importantly, the distribution of PNs

311

for each cell was similar for both stretched and unstretched tissue (Fig. 3C), showing that

312

GAT can capture the underlying cellular distribution even with significant differences in

313

tissue area.

314

Spatial analysis in GAT enables the quantification of cellular distribution and can be used to

315

evaluate differences across gut regions or in different disease states. To illustrate this

316

objectively, images of the proximal colon (PC) and distal colon (DC) were analyzed from a

317

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

published dataset (Hamnett, Dershowitz, Gomez-Frittelli, et al. 2022; Hamnett, Dershowitz,

318

Sampathkumar, et al. 2022). Others have demonstrated that the PC has larger ganglia and

319

greater neuronal density in comparison to the DC (Li et al. 2019; Nestor-Kalinoski et al.

320

2022). This was confirmed by our analysis using GAT (neurons per ganglion: 20.7 ± 30.3 for

321

DC vs 56.5 ± 202.3 for PC, Mean ± S.D. respectively, unpaired t-test with Welch's

322

correction, p<0.0001, n=4). The large S.D. values are reflective of the variations in ganglia

323

sizes for each region. Neurons in the PC had greater number of neighbors than those in the

324

DC (Fig. 3D; raw counts: 503.8 ± 68.6 vs. 874.8 ± 161 for 5 neighbors, 180.3 ± 20.4 vs.

325

451.5 ± 120.2 for 6 neighbors for DC vs. PC respectively, Mean ± S.D., unpaired t-test with

326

Welch's correction, p<0.01, n=4). This supports the observation of larger neuronal counts per

327

ganglion in the DC. However, to account for differences in neuronal numbers in the PC and

328

DC, the raw neighbor counts were normalized to total neuron count for each region. This

329

revealed significant regional differences and a shift in the neuronal distribution (Fig. 3E). In

330

the DC, a significant proportion of neurons had fewer neighbors, which is indicative of

331

smaller ganglion size and, consequently, smaller neuronal clusters. The larger ganglion size

332

and neuronal clusters were evident from the larger proportion of cells having more neighbors

333

(Fig. 3E and F) in the PC relative to the DC (normalized counts: 0.27 ± 0.01 vs. 0.23 ± 0.01

334

for 3 neighbors, 0.14 ± 0.005 vs. 0.2 ± 0.007 for 5 neighbors, 0.05 ± 0.004 vs. 0.1 ± 0.013 for

335

6 neighbors for DC vs. PC respectively, Mean ± S.D., unpaired t-test with Welch's correction,

336

p<0.001, p<0.0001 and p<0.01 respectively, n=4). Thus, the number of PNs is proportional to

337

the cell density and size of the ganglion.

338

Region-specific differences in the distribution of neuronal subtypes could be reflective of the

339

specific functions of the GI subregions (Hamnett, Dershowitz, Sampathkumar, et al. 2022; Li

340

et al. 2019; Nestor-Kalinoski et al. 2022). The regional distribution of the calcium-binding

341

proteins, Calb+ and Calret+ neurons of the PC and DC were examined using images from the

342

Hamnett et al. (2022) dataset. The aim was to test the capacity of GAT to detect established

343

regional differences in ENS distribution and investigate the relative distribution of these

344

neuronal markers as they were co-labeled in the same tissue (Fig. 3G). Using GAT analysis,

345

a higher proportion of Calb+ neurons were detected in the PC compared to the DC (Fig. 3H;

346

1.99 ± 0.78 for DC vs 6.25 ± 2.71 for PC, Mean ± S.D., unpaired t-test with Welch's

347

correction, p<0.05, n=4). Conversely, no significant differences in the number of Calret+

348

neurons were detected across regions (Fig. 3I; 26.96 ± 4.9 for DC vs 23.51 ± 9.3 for PC,

349

Mean ± S.D., unpaired t-test with Welch's correction, p>0.5, n=4) consistent with Hamnett et

350

al. (2022). In this dataset, preparations were co-labeled for Calb and Calret, enabling the use

351

of GAT to assess neurons that were double positive for these markers. A very small

352

proportion of Calret and Calb double positive neurons were detected in PC and DC, and no

353

significant difference was found in the distribution across regions (Fig. 3J; 0.65 ± 0.46 for

354

DC vs 1.54 ± 0.71 for PC, Mean ± S.D., unpaired t-test with Welch's correction, p>0.05,

355

n=4).

356

The distribution of Calb+ and Calret+ neurons was further assessed using spatial analysis in

357

GAT. This revealed significantly greater numbers of neurons with one Calb+ neighbor in the

358

PC compared to the DC (Fig. 3K; 0.18 ± 0.06 for PC vs. 0.06 ± 0.02 for DC, Mean ± S.D.,

359

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

unpaired t-test with Welch's correction, p<0.05, n=4). This finding aligns with the

360

observation that the PC contains a greater proportion of Calb+ neurons (Fig. 3H), resulting in

361

a larger number of Calb+ neighbors surrounding each neuron compared to DC. No difference

362

between regions was detected for neurons with Calret+ neighbors (Fig. 3I), which aligns with

363

the finding of lower proportion of Calret+ neighbors in both DC and PC. The spatial

364

distribution of neurons that coexpressed Calb+ and Calret+ was also determined, with no

365

significant difference between regions (Fig. 3M-N). Spatial analysis of Calret+ neurons

366

relative to Calb+ neurons and vice versa did not reveal any regional differences, suggesting

367

that Calret+ neurons do not preferentially associate with Calb+ neurons in either the PC or

368

DC. These analyses demonstrate that spatial analysis using GAT effectively detects regional

369

differences in neuronal and neuronal subtype distribution.

370

Discussion

371

GAT is a user-friendly Fiji-based software for studying the distribution of enteric neurons

372

and their neurochemical coding in whole mounts of GI tissue in 2D. It uses DL models for

373

segmentation of neurons and ganglia which enables higher throughput and faster data

374

extraction making it possible to analyze large tissue areas. Moreover, spatial analysis in GAT

375

provides an objective means for characterizing the distribution of cells and distinct

376

subpopulations. The DL models are coupled with a user-friendly workflow thus enabling

377

researchers with minimal computational expertise to adopt GAT for rapid and reproducible

378

image analysis.

379

The availability of state-of-the-art user-friendly tools such as StarDist (Schmidt et al. 2018)

380

and deepImageJ (Gómez-de-Mariscal et al. 2021), in combination with ZeroCostDL4Mic

381

(von Chamier et al. 2021) was crucial for the development of GAT. The GAT enteric neuron

382

models can be used within any software that supports StarDist, thus giving the user flexibility

383

to generate custom analysis pipelines should they be required. The GAT software repository

384

has models and scripts compatible with QuPath, a popular image analysis software for

385

analyzing whole slide images and large 2D images. Cellpose was used to compare the

386

segmentation abilities of the DL models as it works readily on diverse datasets and the user

387

interface makes it simple to use. It is unclear if the cyto2 model was trained on enteric

388

neuronal datasets, which may explain the slightly lower performance metrics when compared

389

to GAT (Fig. 2). In Cellpose v2.0, additional models trained on fluorescent images are

390

available, which may show an improvement over cyto2 (Pachitariu & Stringer 2022).

391

Moreover, the best performance can be achieved by fine-tuning the Cellpose models, where

392

the GAT training dataset can be combined with user-specific data to generate custom models

393

(Pachitariu & Stringer 2022). This could increase the accuracy of the segmentation. However,

394

performing quality checks with objective metrics is essential, as shown in Fig. 2A, 2C, and

395

S6.

396

StarDist was used for segmenting cells within GAT, as the software is optimized to detect

397

objects with star-convex shapes such as cell nuclei. Thus, it is suitable for detecting enteric

398

neurons that have a circular shape (Schmidt et al. 2018). Other important benefits include: the

399

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

availability as a Fiji plugin; the ability to use StarDist within macros/scripts; the ability to

400

tune the cell detection by changing the “probability” value, allowing segmentation of images

401

with varying labelling intensities or with high background noise and the ability to detect

402

overlapping cells. The overlap detection can be adjusted by changing the “overlap threshold”

403

value. This is particularly useful for analyzing tissue that has not been stretched appropriately

404

or tissue from larger animals where the ganglia are thicker resulting in greater overlap

405

between cells in 2D. The caveat of using StarDist is that it can only be used for round cells

406

and not for cells with complex morphology. This currently limits its use to enteric neurons,

407

and other celltypes where a nuclear stain is available, such as Sox10 for enteric glia. Thus,

408

other cells with non-circular complex shapes, such as tissue resident macrophages or

409

interstitial cells of Cajal, cannot be analyzed using the current pipeline. Future versions of

410

GAT aim to add support for analyzing diverse cell types within the gut wall.

411

A limitation of using DL-based models in GAT is that they may not work across image types

412

that GAT has not previously encountered. This could be rectified by retraining the models

413

with new data, but it may not always be feasible as this process is laborious and requires

414

computational expertise. Given the evolving landscape of image analysis and cell

415

segmentation software, GAT offers an option of importing custom segmentations for cells

416

and ganglia directly into the analysis pipeline. This feature allows flexibility for the user to

417

choose their preferred cell segmentation tool. As an example this approach was used in Fig.

418

3. The neuron subtype model successfully segmented Calb+ neurons but was not consistent

419

for Calret+ neurons as the labelling was heterogenous. Similarly, ganglia segmentation was

420

not consistent using the ganglia model. To rectify this, QuPath was used for training an object

421

classifier for Calret+ neurons and a pixel classifier for ganglia, thus enabling segmentation of

422

ganglia and Calret+ neurons respectively. The respective detections and annotations were

423

exported from QuPath and reimported back into GAT during analysis.

424

One limitation of GAT is that the analysis workflow is currently designed for 2D images, as

425

the Fiji StarDist plugin (v0.3.0) is limited to 2D datasets. Currently, GAT does not support

426

importing 3D segmentation from other software. As cells which occupy a 3D space are

427

projected onto a 2D image, they can often superimpose or overlap with each other, leading to

428

challenges in accurately delineating and segmenting these cells. Separation of cells is more

429

readily achieved with 3D data; however, only the Python implementation of StarDist supports

430

3D segmentation. Furthermore, cell shape and size are more accurately measured in 3D

431

compared to 2D. One reason is that the volume measurements are less impacted by

432

differences in tissue stretch compared to area measurements in 2D. 3D segmentation requires

433

DL-based approaches that use high-quality 3D annotations, which is a time-consuming and

434

laborious process. Several tools are available for annotating data in 3D (Berg et al. 2019;

435

Boergens et al. 2017; Borland et al. 2021; Fedorov et al. 2012; Tasnadi et al. 2020). However,

436

the success of DL models relies on the availability of large amounts of high-quality training

437

data, which account for the diversity of the markers used, animal species studied, how the

438

tissue is prepared, and the variability in the instruments used for acquisition. Existing tools

439

such as Cellpose (Stringer et al. 2021), 3D ImageJ suite (Ollion et al. 2013), and CLIJ or

440

pyclesperanto (Haase, Jain, et al. 2020; Haase et al. 2022; Haase, Royer, et al. 2020) can be

441

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

used for processing 3D data and enabling curation of labeled data. To enable ENS-specific

442

analytical solutions, it is essential to have robust training data made accessible to the wider

443

research community. For example, the GAT training dataset was deposited on Zenodo, an

444

open data repository. This has been utilized to generate custom cell segmentation models for

445

image analysis software to quantify Ca

signaling in the gut (Barth et al. 2023).

446

Manual analytical approaches to assess changes in the ENS are limited to cellular density and

447

the number of neuron subtypes within the tissue. However, incorporating spatial analysis can

448

reveal insights into cellular distribution, interactions, and potential implications for function

449

at a tissue-level (Nestor-Kalinoski et al. 2022). Existing spatial analysis software solutions

450

require significant computational expertise and may require optimization to study the ENS

451

(Feng et al. 2023; Rose et al. 2020; Stoltzfus et al. 2020; Yang et al. 2020). Factors such as

452

operator expertise, animal species being studied, and the pathology investigated can affect

453

how the tissue is stretched and prepared as a wholemount (Gomez-Frittelli, Hamnett &

454

Kaltschmidt 2023; Kapur 2013). This, in turn, can affect analysis and interpretation of

455

cellular distribution (Fig. 3A-C). GAT accounts for possible variations in tissue stretch by

456

using an enteric neuron-specific nearest neighbor distance threshold value. Differences in

457

neuronal distribution were demonstrated objectively using the proximal neighbor analysis in

458

GAT. The PC had relatively larger neuronal clusters compared to the DC indicative of the

459

large and small average ganglia sizes, respectively. Altered distribution of Calret+ and Calb+

460

neurons across the PC and DC were reflected in the spatial analysis, and no spatial

461

association was found between neuronal subtypes. This approach can be extended to assess

462

changes in the structure or morphology of the ENS across gut regions or in diseases such as

463

inflammatory bowel disease (Brierley & Linden 2014), diabetes (Chandrasekharan et al.

464

2011; Ingrid et al. 2013) and enteric neuropathies, such as Hirschsprung disease (Heuckeroth

465

2018). For example, this analysis could be used to determine if there are subtle differences,

466

such as fewer nNOS+ neuronal neighbors, which would be indicative of lower nNOS+

467

neurons in these conditions. The parameters, such as the average PN distance, may need to be

468

modified for human tissue, and the option to test different values is available in GAT. Due to

469

the analysis capabilities of GAT being limited to enteric neurons, the spatial analysis does not

470

capture the complexity of cellular distribution and relationships with other cell types, such as

471

enteric glia, macrophages, and interstitial cells of Cajal. Incorporating additional spatial

472

analysis metrics such as those related to cell colocalization, spatial heterogeneity (Feng et al.

473

2023), or the use of a spatial neighbors graph to understand neighborhood enrichment (Palla

474

et al. 2022) could enable a more comprehensive and unbiased examination of cellular

475

interactions and distributions in the gut.

476

GAT excels at segmenting neurons and works across images of varying staining qualities. It

477

offers a faster alternative to manual analysis and has been designed with ENS-specific

478

analysis solutions, such as studying neurochemical coding and spatial analysis of neuronal

479

distribution. New features are regularly being introduced based on user feedback and

480

experience. To enable compatibility for further analysis in other software, segmentation

481

maps, cell outlines, cell type information, and cell coordinates are extracted for each

482

experiment during analysis. GAT is written in Fiji using the macro language, limiting the

483

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

scope of the user interface (UI) and its interactivity. Future versions of GAT may use

484

scripting languages in Fiji, which offers more flexibility in developing highly interactive UIs.

485

Moreover, the availability of StarDist models enables the development of interactive

486

workflows in napari and QuPath. GAT aims to create a toolbox that automates common

487

image analysis tasks in ENS research, eliminating the burden of manual analysis. This allows

488

scientists to spend more time interpreting biology and advancing scientific research.

489

Methods

490

Mice: The details on housing conditions and ethics for the mice used within this study can be

491

found in the respective studies (DiCello et al. 2020; Hamnett, Dershowitz, Sampathkumar, et

492

al. 2022; McQuade et al. 2021; Nestor-Kalinoski et al. 2022; Poon et al. 2022).

493

B6;129S6-Gt(ROSA)26Sor

tm9(CAG-tdTomato)Hze

/J (Jackson Laboratory, Bar Harbor, ME; stock

494

no. 007905) mice were crossbred with B6.Cg-E2f1

Tg(Wnt1-cre)2Sor

/J (Jackson Laboratory; stock

495

no. 022501) mice. The F1 offspring expressed tdTomato within all ENS cells. The animals

496

were handled in accordance with the institutional guidelines of the University of Tübingen,

497

which conform to international guidelines. Mice were housed in standard plastic cages with

498

standard bedding under a 12-hour light-to-dark cycle at 22 ± 2

C, 60% ± 5% humidity, with

499

free access to food and water.

500

Rat:

The details on housing conditions and ethics for the rats can be found in (Furness et al.

501

2023).

502

Human Tissue: Pediatric colon tissue was obtained from 1 patient, (male, 4months old with

503

rectosigmoid disease with prior written informed consent (Royal Children’s Hospital and

504

Monash University (HREC: 38262 and 21091)). The ethics for the adult human tissue can be

505

found in Chen et al. (2023).

506

Whole mount tissue preparation and immunohistochemistry:

507

Mouse: The protocols used for dissection and immunohistochemistry of mouse myenteric

508

wholemounts varied across labs and are summarized in the respective publications from each

509

source (DiCello et al. 2020; Hamnett, Dershowitz, Sampathkumar, et al. 2022; McQuade et

510

al. 2021; Nestor-Kalinoski et al. 2022; Poon et al. 2022).

511

Human: The protocol for human colon samples can be found in Chen et al. (2023).

512

Table 1. Primary antibodies used for immunofluorescent labeling of wholemount preparations.

513

Target Host

Species

Dilution Supplier Clone/Catalog

number

RRID/ Reference

HuC/D

Mouse

1/200

Thermo
Fisher

A-21271

AB_221448

HuC/D Human

1:5000

Gift

from

Vanda
Lennon,

N/A AB_2314657

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Mayo Clinic

nNOS Sheep

1:1000

Emson K205

AB_2895154

Calbindin Rabbit

1:1600 Swant

CB-38a

AB_10000340

NFM Rabbit 1:500

Merck AB1987 AB_91201

Calretinin Goat

1:1000 Swant

CG1

AB_10000342

nNOS

Sheep

1:1000

Gift from Dr
Piers Emson

AB_2314960

The secondary antibodies used were donkey anti-human 594 (Jackson ImmunoResearch),

514

donkey anti-sheep 647 and donkey anti-sheep 488 (Thermo Fisher Scientific). Streptavidin-

515

AMCA was used with donkey anti-mouse biotin for Hu in the adult human colon

516

preparations (Chen et al. 2023).

517

External Datasets for model training:

518

External datasets were also used for creating DL models. Data from the Stimulating

519

Peripheral Activity to Relieve Conditions (SPARC) program website (https://sparc.science)

520

were used for training the neuronal, neuronal subset and ganglia segmentation models.

521

Within the SPARC portal, data from mouse (Nestor-Kalinoski et al. 2022; Wang et al. 2021)

522

and human myenteric plexus whole mounts (Chen et al. 2023; Graham et al. 2020b) were

523

used for curating training datasets.

524

Image acquisition

525

Datasets generated for this study were acquired using different instruments. This enables the

526

DL models to generalize well and work across a variety of data acquired with commonly

527

used microscopes.

528

The images for the mouse tissue were acquired using:

529

•

Leica TCS-SP8 confocal system (20x HC PL APO NA 1.33, 40 x HC PL APO NA

530

1.3)

531

•

Leica TCS-SP8 Lightning confocal system (20x HC PL APO NA 0.88)

532

•

Zeiss Axio Imager M2 (20X HC PL APO NA 0.3)

533

•

Zeiss Axio Imager Z1 (10X HC PL APO NA 0.45)

534

Human tissue images were acquired using:

535

•

IX71 Olympus microscope (10X HC PL APO NA 0.3)(Chen et al. 2023)

536

•

Leica TCS-SP8 confocal system (20x HC PL APO NA 1.33, 40 x HC PL APO NA

537

1.3)

538

Acquisition conditions for SPARC datasets are:

539

•

Leica TCS SP5 laser scanning confocal microscope (20x NA 0.70, 40x NA 1.25, or

540

63x NA 1.4) (Nestor-Kalinoski et al. 2022)

541

•

Zeiss LSM 710 confocal microscope (10x and 20x PL APO). Z-axis increments were

542

m (10x objective) or 1

m (20x objective) (Graham et al. 2020b).

543

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Software

544

Training data

545

Training data were generated using custom scripts written in Fiji macro language, provided

546

with GAT (GAT -> Tools -> Data Curation). Briefly, the entire image or a portion of the

547

image was selected for annotation. This was followed by manually outlining neurons using

548

the drawing tools and saving them in the ROI Manager. For the enteric neuron datasets, the

549

annotated images of varying sizes were saved as raw images and as segmented label images,

550

where each neuron had an individual pixel value. In the label images all pixels with value 1

551

belong to neuron 1, all pixels with value 2 belong to neuron 2 and so on. The binary masks

552

were saved for ganglia segmentation. The annotation was performed and verified by at least

553

two researchers experienced with the identification of enteric neurons and ganglia. Briefly,

554

the outlines from the label image were overlaid on the raw images and verified for each cell

555

inusing Fiji. The “Verify Images Masks” macro within GAT -> Tools -> Data Curation were

556

also used. For images that had DAPI labelling, the neuronal nuclei were used to delineate

557

overlapping cells. Incorrect ROIs were deleted and redrawn using the Oval or Freehand

558

drawing tools in Fiji.

559

Enteric neuron models

560

The training images (Tables S1 and S2) were normalized to account for the any variations in

561

the sizes of cells in pixels due to image acquisition conditions, such as resolution and

562

magnification, and species differences. The images from mouse and rat tissue were scaled to

563

a pixel size of 0.568 µm per pixel, whereas the images from human tissue were scaled to 0.9

564

µm per pixel due to the larger cell sizes. This rescaling process ensured that the training

565

images had a uniform average neuron area of 701.2 ± 195.9 pixel

(Mean ± SD, 6267 cells)

566

irrespective of image magnification or animal species. A similar approach was used to

567

generate a training dataset for the enteric neuron subtype model. This training dataset

568

contained images of neurons expressing various neurochemical markers, including Calb,

569

nNOS, CalR, choline acetyltransferase (ChAT), delta-opioid receptor (DOR), mu-opioid

570

receptor (MOR), neurofilament 200 (NF200) and somatostatin. The average cell area in the

571

neuronal subtype dataset was 880.9 ± 316 pixel

(Mean ± SD, 924 cells), with around 56.6%

572

of the cells being nNOS positive. Thus, nNOS cells were overrepresented in the dataset. The

573

StarDist v0.3.0 Fiji plugin was used for inference.

574

Ganglia model

575

The ganglia model was trained on images (Table S3) with both the pan-neuronal marker, Hu

576

and a second marker that labeled the neuronal or glial fibers. Regions where both markers

577

were co-distributed were manually labeled as ganglia. The markers used for the identification

578

of ganglionic structures consisted of any of the following: Protein Gene Product 9.5

579

(PGP9.5), nNOS, glial fibrillary acid protein (GFAP), S100b, Tuj1, or NF200. The

580

DeepImageJ v2.1.12 Fiji plugin was used for inference.

581

Deep learning models and software

582

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

StarDist v0.7.3 (Schmidt et al. 2018) was used via ZeroCostDL4Mic v1.13 notebooks (von

583

Chamier et al. 2021) within Google Colab for training the 2D segmentation models for

584

enteric neurons and neuronal subsets. The ganglia model was trained in Google Colab using a

585

2D UNet architecture (Ronneberger, Fischer & Brox 2015) and exported to be readily used

586

within deepImageJ (Gómez-de-Mariscal et al. 2021). The notebooks used for training the

587

models, the training datasets used, training reports, model quality reports, and the models are

588

deposited online at:

https://doi.org/10.5281/zenodo.6096664.

589

Cellpose (v0.7) has been used as a baseline for comparing cell segmentation in this study. It

590

is a generalist cell segmentation solution aimed at analyzing a wide variety of cell types

591

(Stringer et al. 2021). It is not known if Cellpose has been trained on enteric neuron images.

592

All training data and deep learning models are deposited on Zenodo:

593

https://zenodo.org/doi/10.5281/zenodo.6094887

594

595

COUNTEN analysis

596

For benchmarking with COUNTEN (Kobayashi et al. 2021), the software was accessed from:

597

https://github.com/KLab-JHU/COUNTEN

. The analysis used the default values of sigma and

598

the minimum number of neurons per ganglia, set at 7 and 3 respectively. An Google Colab

599

notebook was designed to enable interactive analysis with an option for batch analysis, which

600

can be accessed from:

https://github.com/pr4deepr/COUNTEN

601

602

Analysis of calbindin and calretinin positive cells

603

A publicly available dataset (Hamnett, Dershowitz, Gomez-Frittelli, et al. 2022) was used for

604

analyzing Calb and CalR-positive cells. Image files with Calb and CalR co-labelling (EXP

605

174) were analyzed using a combination of QuPath v0.4.3 (Bankhead et al. 2017) and GAT.

606

Due to inconsistent segmentation of the ganglia using the ganglia model, a pixel classifier

607

was trained in QuPath to identify ganglia based on the co-expression of Hu, Calb and CalR.

608

The resulting annotations were exported from QuPath as Fiji-compatible ROIs, which were

609

subsequently imported into GAT for analysis. Similarly, the enteric neuron subtype model

610

detected CalB-positive neurons but did not reliably detect CalR-positive cells. To address

611

this, an object classifier was trained in QuPath to detect CalR-positive neurons. The neurons

612

were initially detected using the enteric neuron StarDist model based on the Hu channel,

613

followed by the application of the object classifier. Only the calretinin-positive ROIs were

614

extracted from QuPath and imported into GAT for analysis.

615

The results for each replicate were merged into summary data using the scripts within GAT -

616

> Analysis. The summary data were analyzed in Python (v 3.9.15) and pandas (v 2.0.2) and

617

visualized using seaborn (v 0.12.2). The analyzed data were exported, and statistical analysis

618

was performed in GraphPad Prism (v 9.5.1).

619

Other

620

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

ChatGPT (GPT-3.5) was used for initial formatting and editing of the manuscript.

621

622

623

Supplementary

624

625

626

627

Figure S1: Effects of sample preparation and tissue stretch on the ENS. For wholemount

628

preparations, the isolated intestine is cut open along the mesenteric border, stretched, and pinned as a

629

flat sheet. Individual expertise can determine the quality of this dissection and preparation (Some

630

parts of the image were created with Biorender.com).

631

632

633

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

634

Figure S2. Neuronal density estimates are highly variable at lower magnifications and sampling:

635

(A) An image of mouse colon tissue labeled for the pan-neuronal marker Hu (cell density 13.9 mm

636

area). Overlays represent the area occupied by different magnifications (10X, 20X, and 40X). (B)

637

Using a StarDist-trained model, QuPath was used to segment 6650 neurons (blue overlay). Regions

638

with inconsistent staining or dissection were excluded (yellow overlay). Each yellow tile is

639

representative of a 40X field of view (FOV) (C) For each individual tile, the density of neurons was

640

calculated to demonstrate that 20X and 40X had a larger variation compared to 10X. (D) To illustrate

641

the effect of sampling on averaged neuronal density estimates, images from each magnification were

642

randomly sampled and mean cell densities were estimated across different sampling rates (2-10

643

images). This was repeated 100 times for each FOV and sampling number. Each bar in the graph

644

represents mean estimates of mean densities for each sampling and FOV choice. Bars represent the

645

minimum and maximum value. The black dotted line indicates the cell density calculated from the

646

entire field of view in A, 0.4768 (x 0.001 cells/µm

647

648

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

649

Figure S3: Random sampling of tiles to determine mean cell counts for different FOVs.

650

An illustration for Fig. S2D outlining how calculations were performed. The image with Hu

651

labelling was divided into tiles for each FOV (10X, 20X, or 40X). To simulate experiments

652

for a specific magnification with different sample numbers, a range of samples from 2 to 10

653

was chosen. The mean cell density was calculated for each sample number, and this was

654

repeated 100 times to estimate the range of mean values that could be obtained.

655

656

Figure S4: Training and validation loss for different models. Training loss of the StarDist

657

enteric neuron model (A) and enteric neuron subtype model (B).

658

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

659

660

Figure S5: Comparison of StarDist segmentations. Comparison between representative

661

StarDist (cyan outline/text) segmentations and manual annotation (yellow outline). The

662

StarDist predictions have smoother boundaries, but do not capture the original cell contours

663

(Scale bar = 10 µm).

664

665

Figure S6: Evaluation of Cellpose and StarDist segmentation metrics for enteric

666

neurons. Cellpose had a lower Adapted Rand Error indicating higher overall segmentation

667

quality than the StarDist model (A). However, this could be because Cellpose is better at

668

predicting accurate cell borders compared to how StarDist approximates the cell shape using

669

star convex polygons. The Cellpose model had larger merge and split errors (B & C) which

670

may explain the higher percentage cell count error presented in Figure 2.

671

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

672

Figure S7: Average proximal neighbor distances can be determined by calculating the

673

local thickness within the ganglia. The distances between neurons in a ganglion can be

674

determined by segmenting the neurons and the ganglia (A, B) and measuring the ganglionic

675

area, excluding neuronal soma (C). (D) The local thickness map can be used to derive the

676

interneuronal distances. This was calculated by applying the “Local Thickness” plugin

677

(Dougherty & Kunzelmann 2007) on the binary image (C). The “Local Thickness (complete

678

process)” option in Fiji was used to compute the local thickness map as shown in D.

679

680

Table S1: Summary of the images used for training the enteric neuron StarDist model. All

681

images contain cells labeled with the pan-neuronal maker Hu. The number of images are mentioned

682

under the species and Total images.

683

Species

Lab 1

Lab 2

Lab 3

SPARC

Lab 4

Total

images

Mouse

Rat

2 - - -

Human

Magnification

40X/20X 10X

20X

20X/40X 10X

Modality

Confocal Widefield Widefield Confocal Widefield

Total images

684

Table S2: Number of images used for training the enteric neuron subtype StarDist model.

685

Data source

nNOS

MOR DOR ChAT CalR CalR NFM CGRP SST Total/

source

Lab 1

4 2 1 2 1

- - - - 10

Lab 2

16 - - -

- - - 16

Lab 3

3 - - - 5

9 6 - - 23

SPARC

1 - - 3

1 - 1 1 7

Total/
marker

24 2 1 5

10 6 1 1 56

686

687

688

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Table S3: Number of images used for training the ganglia model: Each marker was used in

689

combination with Hu to generate training data for the 2D U-Net ganglia model implemented using

690

deepImageJ.

691

Data
source

PGP9.5 GFAP Wnt1Cre-

GFP

ChAT NF200 5HT CGRP NPY nNOS s100b Tuj1 Total

/
source

Lab 1

12 28 3

- - - - - 5 - - 48

Lab 2

Lab 3

- - -

- -

- - -

Lab 4

SPARC - - -

4 - -

- -

- 3

Total/
marker

12 28

5 1 2

1 1

3 3

692

Supplementary methods

693

Effects of varying magnification and sampling

694

An image of a myenteric wholemount of the mouse colon (13.9 mm

) labeled with Hu was

695

used with QuPath v0.3.2 (Bankhead et al. 2017) to test the effects of varying magnification

696

and sampling on estimated cell counts (Figs 2 & 3). A whole image annotation was created

697

and then divided into tiles, where each tile had areas of 775,918 µm

, 338,116 µm

and

698

150,274 µm

, thus simulating 10X, 20X, and 40X magnifications, respectively. Tiles at the

699

edges of the tissue that were below 60,000 µm

in area and areas with uneven staining were

700

excluded in the subsequent calculations. To perform cell segmentation, a custom groovy

701

script in combination with the StarDist 2D enteric neuron model converted into ONNX

702

format was used in QuPath

703

(

https://github.com/pr4deepr/GutAnalysisToolbox/tree/main/QuPath_workflow

). The

704

parameters used for segmentation were a rescaling factor of 1 and a probability of 0.65. Once

705

segmentation was performed, the cell numbers were saved using the annotation measurement

706

tables. Once cell counts were estimated, a custom Python script was used to choose random

707

tiles and estimate mean values. This is illustrated in Figure S1. The Python code is provided

708

as a Jupyter notebook.

709

710

Evaluation of CellPose and StarDist segmentation

711

The segmentation metrics for Cellpose cyto2 and StarDist enteric neuron model were further

712

evaluated using Adapted Rand error, VOI merge and VOI split on the test data from the

713

enteric neuron model training. Adapted Rand error assesses the overall segmentation quality,

714

whereas VOI merge and VOI split assess errors associated with cell merging and splitting

715

respectively (Wolny et al. 2020). The evaluation script from plant-seg-tools GitHub

716

repository (

https://github.com/hci-unihd/plant-seg-tools/tree/main

) was used to evaluate the

717

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

segmentation results from the Cellpose cyto2 and enteric neuron StarDist model against the

718

ground truth data (Fig. S6).

719

Supplementary text

720

Effects of sample preparation and tissue stretch on neuronal counts

721

Gastrointestinal wholemount preparations are typically used to quantify enteric neuron

722

numbers. More accurate estimates can be achieved with these preparations in comparison to

723

tissue sections, as all neurons within the same plane as the plexuses can be visualized within

724

the 3D or 2D space, and larger tissue samples can be prepared to increase sampling. To

725

prepare a wholemount, the tissue is cut open along the mesentery, stretched and pinned for

726

further processing (illustrated in Fig. S1A). This flat sheet preparation allows investigation of

727

the anatomical features across larger tissue areas. Depending on the expertise of the

728

researcher, inconsistent stretching and uneven sampling can lead to biased estimation of cell

729

numbers (Fig. 3A). One way to mitigate this effect of stretching is to compare the dimensions

730

of the tissue before and after stretching and include a correction factor for changes in tissue

731

dimensions caused by stretching and any age-associated changes (Gamage et al. 2013).

732

However, very few studies have used this approach (Gamage et al. 2013; Kapur 2013; Michel

733

et al. 2022b).

734

735

Effects of inadequate sampling and image acquisition

736

Advances in imaging technologies and software have enabled the sampling of large tissue

737

areas as tile scan images, a standard feature in many conventional widefield and confocal

738

microscopes. However, the analysis throughput of large images remains a significant limiting

739

factor.

740

741

Due to the time-consuming nature of the manual analysis, the quantification of neurons along

742

larger lengths of the intestine is not standard practice within the field (Michel et al. 2022b;

743

Nestor-Kalinoski et al. 2022). Estimates of neuron densities are calculated from randomly

744

chosen FOVs selected from small sections of the gut, which are averaged. These values are

745

representative of total numbers across entire regions of the intestine. The following section

746

aims to quantify the errors associated with FOV selection and choice of magnification during

747

image acquisition.

748

749

A large mouse proximal colon myenteric wholemount (13.9 mm

) was labeled for Hu (Fig.

750

S2A). GAT StarDist model converted to ONNX format for QuPath was used to calculate a

751

cell density of 0.4768 (x 0.001 cells/µm

) across the tissue. To simulate the effects of

752

sampling using single images at 10X, 20X, and 40X, the whole image was divided into

753

smaller tiles equating to these magnifications on a Leica SP8 confocal microscope (Figs. S2A

754

& B). The cell count was estimated for each tile and plotted as a frequency histogram (Fig.

755

S2C). The density values for 20X and 40X FOVs had a larger variability compared to 10X

756

(Range (max – min): 0.3844 for 10X, 0.5766 for 20X and 0.5913 for 40X). Thus, there is

757

increased variability when estimating cell density with higher magnifications, as only smaller

758

tissue areas can be sampled.

759

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

760

The choice of magnification and sample number could affect the accuracy of the final cell

761

counts. Cell counts were averaged from 2 to 10 tiles randomly selected at each magnification

762

to measure the effects of sample size and subsequently averaged (Fig. S2C). To estimate the

763

range of mean values that could be obtained, this process was repeated 100 times (Fig. S3).

764

Each bar in Fig. S2D shows 100 possible mean density estimates for all the different sample

765

numbers. The larger the bar, the higher the variability in the final estimate. The actual cell

766

density is calculated for the entire FOV and is denoted by the black dotted line. When

767

comparing 10X and 40X FOVs, 10 images of the 40X FOV were required to have a

768

comparable variability to 5 images from 10X FOV (Range of 0.182 for 10X and 0.27 for 40X

769

for 5 samples vs Range of 0.1 for 10X and 0.164 for 40X for 10 samples). Thus, it is

770

recommended to use larger FOVs, such as a 10X or 20X magnification and increased tissue

771

sampling to ensure accurate cell density estimates.

772

773

Extensibility of GAT models

774

The StarDist models are compatible with other image analysis software, including QuPath

775

(Bankhead et al. 2017) and napari (Sofroniew et al. 2022). This compatibility enables users to

776

create custom analysis pipelines without being restricted to a single software platform. Some

777

examples include:

778

•

QuPath is a popular software for analyzing large 2D images that has an extension for

779

implementing StarDist cell segmentation (Schmidt et al. 2018). The 2D enteric neuron

780

models have been converted into the ONNX format using tf2onnx software

781

(

https://github.com/onnx/tensorflow-onnx)

to make it accessible from QuPath. A

782

groovy script and the QuPath model are provided in the GAT repository. An

783

advantage of QuPath is its ability to navigate large 2D images easily and use less

784

computing power than Fiji when performing StarDist segmentation on large tile scans.

785

•

napari is a popular Python-based multi-dimensional viewer. The availability of an

786

interactive napari StarDist plugin (

https://github.com/stardist/stardist-napari

) enables

787

deployment of the StarDist enteric neuron models within a Python image processing

788

workflow. Moreover, as the ganglia model is a TensorFlow model, it can be

789

implemented in Python along with the StarDist models.

790

•

APEER is a cloud-based image processing platform by Zeiss that allows custom

791

image processing workflows to be run directly from the image acquisition software

792

Zen. It has a StarDist module (

https://www.apeer.com/app/modules/StarDist-2D-

793

Nucleus-Segmentation-%28beta%29/52cc57d2-94ea-4417-a4c9-d8d957a3cf32

). The

794

GAT StarDist models can likely be used within Zen. Namely, GAT workflows could

795

be set up to perform cell counts immediately after image acquisition.

796

•

CellProfiler is a popular cell image analysis software that does not require prior

797

programming experience (McQuin et al. 2018). It currently has a StarDist plugin

798

(

https://github.com/CellProfiler/CellProfiler-plugins

), which could enable the use of

799

the GAT models within a CellProfiler pipeline.

800

801

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

802

803

Limitation of the DL models

804

The data used for training DL models is crucial for ensuring the model works accurately

805

across a wide range of images. The DL models are only as good as the training data used, and

806

any biases within the training data are carried over within the model (Jacquemet 2021,

807

Litjens, 2017 #72). For example, earlier versions of GAT models trained on 2D maximum

808

projections of confocal images were not accurate on widefield images. Thus, emphasis was

809

placed on gaining diverse training datasets. This does not necessarily guarantee the

810

robustness of the DL models. Instances where DL models may not work well are discussed

811

below.

812

Effect of tissue type, preparation, and species. The training datasets for the models are

813

predominantly from myenteric wholemounts of the mouse colon. As such, the models may

814

not work well on wholemounts of the submucosa, different regions of the gut, or preparations

815

from other species. The GAT enteric neuron models have been tested successfully (not

816

shown) on submucosal regions of the colon, although there are no training data from other

817

layers.

818

Effect of imaging modality and choice of fluorophores: As images from a confocal

819

microscope have different noise levels, contrast, and background labeling compared to

820

widefield microscopy images (Shaw 2006), images acquired with both modalities were

821

included in the training dataset. The training data has 60% confocal 2D image projections and

822

40% widefield images. Moreover, the images consisted of Hu labeling using a range of

823

fluorophores (405, 488, 568, and 647 nm emission). In addition to sample preparation and

824

imaging modality affecting autofluorescence, biological tissue is increasingly autofluorescent

825

at lower wavelengths. Autofluorescence from blood vessels can affect cell segmentation,

826

leading to erroneous segmentation and predictions.

827

Neuronal subtype model: The distribution of labeling in the neuronal soma varies between

828

markers in the training dataset used. This includes punctate staining (delta or mu-opioid

829

receptors), cytoplasmic staining (nNOS), or variable staining intensities (CalR, nNOS). Due

830

to the high variability, the neuronal subtype model may not work across all neuronal markers.

831

This issue was encountered for the analysis in Fig 3. Increasing the size of the dataset with

832

images for different markers or using a different type of neural network architecture may

833

increase the accuracy of the neuronal subtype predictions. Alternatively, using other software

834

such as QuPath or finetuning a Cellpose model on the specific datasets can increase the

835

accuracy of cell detection.

836

Ganglia model: The ganglia segmentation is based on the co-labeling of Hu with another

837

neuronal/glial fiber marker. Within GAT, ganglia are not defined by a minimum number of

838

neurons, so all the cells and ganglion-like structures are included for analysis. This is

839

particularly important in neuropathic conditions such as Hirschsprung disease, where the

840

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

absence and/or relative reduction of enteric neuronal number is used to support its diagnosis

841

(Burns et al. 2016; Heuckeroth 2018; Niesler et al. 2021; Schäppi et al. 2013). The accuracy

842

of the ganglia segmentation model depends on the quality of the neuronal or glial fiber

843

labeling. High background can affect the accuracy of this approach. The DL approach

844

requires two channels (Hu and a ganglia marker) for defining the ganglia, whereas

845

COUNTEN can segment ganglia with just Hu. As the presence of neurons can be used to

846

define the ganglia, GAT has an additional option to customize ganglia detection using only

847

the Hu labelling. Moreover, for challenging cases, segmentation can be performed in QuPath

848

using pixel classification and the custom ROIs imported back into the GAT workflow.

849

The reality of DL-based approaches is that GAT may not work accurately on image types it

850

has not encountered before. Generating accurate models requires creating more annotated

851

data, which is made possible within GAT. It is designed to be part of the routine analysis,

852

where the user can opt in to save the images and masks or can be done separately on smaller

853

FOVs. The resulting data can be used for finetuning the GAT models and can be submitted to

854

a public repository. Making annotated data publicly available enables the generation of

855

domain specific image analysis software that can benefit the wider ENS research community

856

(Barth et al. 2023).

857

858

Additional features in GAT:

859

Calcium Imaging Analysis

860

861

The long-term goal for GAT is to have a collection of common image analysis workflows

862

used within the ENS field. In addition to the analysis of immunofluorescence images, GAT

863

has a calcium imaging analysis workflow aimed at extracting normalized calcium responses

864

from cells in wholemount preparations. The workflow has two parts:

865

866

•

Image alignment. The output of calcium imaging experiments is typically timeseries

867

images which may have movement artefacts either due to drift or tissue movement.

868

GAT provides options to align these images using ‘Linear Stack Alignment with

869

SIFT’ (Lowe 2004) and ‘Template matching’ (Tseng et al. 2011) plugins within Fiji.

870

This can be done on single images or on a folder of images.

871

872

•

Extraction of normalized traces. Once the alignment is performed, the time series

873

image stacks can be imported into Fiji, and the regions of interest (ROIs) manually

874

defined using its drawing tools. The analysis can be performed for multiple cell types

875

in the same image and the extracted traces can be normalized to a user-defined

876

baseline.

877

878

References:

879

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Bankhead, P, Loughrey, MB, Fernández, JA, Dombrowski, Y, McArt, DG, Dunne, PD,

880

McQuaid, S, Gray, RT, Murray, LJ, Coleman, HG, James, JA, Salto-Tellez, M & Hamilton,

881

PW 2017, 'QuPath: Open source software for digital pathology image analysis', Scientific

882

Reports, vol. 7, no. 1, p. 16878.

883

884

Barth, BB, Redington, ER, Gautam, N, Pelot, NA & Grill, WM 2023, 'Calcium image

885

analysis in the moving gut', Neurogastroenterology & Motility, vol. n/a, no. n/a, p. e14678.

886

887

Berg, S, Kutra, D, Kroeger, T, Straehle, CN, Kausler, BX, Haubold, C, Schiegg, M, Ales, J,

888

Beier, T, Rudy, M, Eren, K, Cervantes, JI, Xu, B, Beuttenmueller, F, Wolny, A, Zhang, C,

889

Koethe, U, Hamprecht, FA & Kreshuk, A 2019, 'ilastik: interactive machine learning for

890

(bio)image analysis', Nature Methods, vol. 16, no. 12, pp. 1226-1232.

891

892

Boergens, KM, Berning, M, Bocklisch, T, Bräunlein, D, Drawitsch, F, Frohnhofen, J, Herold,

893

T, Otto, P, Rzepka, N, Werkmeister, T, Werner, D, Wiese, G, Wissler, H & Helmstaedter, M

894

2017, 'webKnossos: efficient online 3D data annotation for connectomics', Nature Methods,

895

vol. 14, no. 7, pp. 691-694.

896

897

Borland, D, McCormick, CM, Patel, NK, Krupa, O, Mory, JT, Beltran, AA, Farah, TM,

898

Escobar-Tomlienovich, CF, Olson, SS, Kim, M, Wu, G & Stein, JL 2021, 'Segmentor: a tool

899

for manual refinement of 3D microscopy annotations', BMC Bioinformatics, vol. 22, no. 1, p.

900

260.

901

902

Brierley, SM & Linden, DR 2014, 'Neuroplasticity and dysfunction after gastrointestinal

903

inflammation', Nat Rev Gastroenterol Hepatol, vol. 11, no. 10, pp. 611-627.

904

905

Burns, AJ, Goldstein, AM, Newgreen, DF, Stamp, L, Schäfer, K-H, Metzger, M, Hotta, R,

906

Young, HM, Andrews, PW, Thapar, N, Belkind-Gerson, J, Bondurand, N, Bornstein, JC,

907

Chan, WY, Cheah, K, Gershon, MD, Heuckeroth, RO, Hofstra, RMW, Just, L, Kapur, RP,

908

King, SK, McCann, CJ, Nagy, N, Ngan, E, Obermayr, F, Pachnis, V, Pasricha, PJ, Sham,

909

MH, Tam, P & Vanden Berghe, P 2016, 'White paper on guidelines concerning enteric

910

nervous system stem cell therapy for enteric neuropathies', Developmental Biology, vol. 417,

911

no. 2, pp. 229-251.

912

913

Caicedo, JC, Roth, J, Goodman, A, Becker, T, Karhohs, KW, Broisin, M, Molnar, C,

914

McQuin, C, Singh, S, Theis, FJ & Carpenter, AE 2019, 'Evaluation of Deep Learning

915

Strategies for Nucleus Segmentation in Fluorescence Images', Cytometry A, vol. 95, no. 9, pp.

916

952-965.

917

918

Cairns, BR, Jevans, B, Chanpong, A, Moulding, D & McCann, CJ 2021, 'Automated

919

computational analysis reveals structural changes in the enteric nervous system of nNOS

920

deficient mice', Sci Rep, vol. 11, no. 1, p. 17189.

921

922

Chandrasekharan, B, Anitha, M, Blatt, R, Shahnavaz, N, Kooby, D, Staley, C, Mwangi, S,

923

Jones, DP, Sitaraman, SV & Srinivasan, S 2011, 'Colonic motor dysfunction in human

924

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

diabetes is associated with enteric neuronal loss and increased oxidative stress',

925

Neurogastroenterology & Motility, vol. 23, no. 2, pp. 131-e126.

926

927

Chen, BN, Humenick, A, Yew, WP, Peterson, RA, Wiklendt, L, Dinning, PG, Spencer, NJ,

928

Wattchow, DA, Costa, M & Brookes, SJH 2023, 'Types of Neurons in the Human Colonic

929

Myenteric Plexus Identified by Multilayer Immunohistochemical Coding', Cell Mol

930

Gastroenterol Hepatol, vol. 16, no. 4, pp. 573-605.

931

932

DiCello, JJ, Carbone, SE, Saito, A, Rajasekhar, P, Ceredig, RA, Pham, V, Valant, C,

933

Christopoulos, A, Veldhuis, NA, Canals, M, Massotte, D & Poole, DP 2020, 'Mu and Delta

934

Opioid Receptors Are Coexpressed and Functionally Interact in the Enteric Nervous System

935

of the Mouse Colon', Cell Mol Gastroenterol Hepatol, vol. 9, no. 3, pp. 465-483.

936

937

Dougherty, RP & Kunzelmann, K-H 2007, 'Computing Local Thickness of 3D Structures

938

with ImageJ', in Microscopy & Microanalysis 2007 Meeting, Ft. Lauderdale, Florida, August

939

5-9, 2007, <

https://www.optinav.info/LocalThicknessEd.pdf

940

941

Fedorov, A, Beichel, R, Kalpathy-Cramer, J, Finet, J, Fillion-Robin, J-C, Pujol, S, Bauer, C,

942

Jennings, D, Fennessy, F, Sonka, M, Buatti, J, Aylward, S, Miller, JV, Pieper, S & Kikinis, R

943

2012, '3D Slicer as an image computing platform for the Quantitative Imaging Network',

944

Magnetic Resonance Imaging, vol. 30, no. 9, pp. 1323-1341.

945

946

Feng, Y, Yang, T, Zhu, J, Li, M, Doyle, M, Ozcoban, V, Bass, GT, Pizzolla, A, Cain, L,

947

Weng, S, Pasam, A, Kocovski, N, Huang, Y-K, Keam, SP, Speed, TP, Neeson, PJ, Pearson,

948

RB, Sandhu, S, Goode, DL & Trigos, AS 2023, 'Spatial analysis with SPIAT and spaSim to

949

characterize and simulate tissue microenvironments', Nature Communications, vol. 14, no. 1,

950

p. 2697.

951

952

Furness, JB 2012, 'The enteric nervous system and neurogastroenterology', Nature Reviews

953

Gastroenterology & Hepatology, vol. 9, no. 5, pp. 286-294.

954

955

Furness, JB, Lei, E, Hunne, B, Adams, CD, Burns, AJ, Wykosky, J, Fazio Coles, TE,

956

Fothergill, LJ, Molero, JC, Pustovit, RV & Stamp, LA 2023, 'Development of the aganglionic

957

colon following surgical rescue in a cell therapy model of Hirschsprung disease in rat',

958

Disease Models & Mechanisms, vol. 16, no. 6, p. dmm050055.

959

960

Gamage, PPKM, Ranson, RN, Patel, BA, Yeoman, MS & Saffrey, MJ 2013, 'Myenteric

961

neuron numbers are maintained in aging mouse distal colon', Neurogastroenterology &

962

Motility, vol. 25, no. 7, pp. e495-e505.

963

964

Gómez-de-Mariscal, E, García-López-de-Haro, C, Ouyang, W, Donati, L, Lundberg, E,

965

Unser, M, Muñoz-Barrutia, A & Sage, D 2021, 'DeepImageJ: A user-friendly environment to

966

run deep learning models in ImageJ', Nature Methods, vol. 18, no. 10, pp. 1192-1195.

967

968

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Gomez-Frittelli, J, Hamnett, R & Kaltschmidt, JA 2023, 'Comparison of wholemount

969

dissection methods for neuronal subtype marker expression in the mouse myenteric plexus',

970

Neurogastroenterol Motil, vol., p. e14693.

971

972

Graham, KD, López, SH, Sengupta, R, Shenoy, A, Schneider, S, Wright, CM, Feldman, M,

973

Furth, E, Valdivieso, F, Lemke, A, Wilkins, BJ, Naji, A, Doolin, EJ, Howard, MJ &

974

Heuckeroth, RO 2020a, Robust 3-Dimensional visualization of human colon enteric nervous

975

system without tissue sectioning, SPARC Consortium, SPARC Consortium.

976

977

Graham, KD, López, SH, Sengupta, R, Shenoy, A, Schneider, S, Wright, CM, Feldman, M,

978

Furth, E, Valdivieso, F, Lemke, A, Wilkins, BJ, Naji, A, Doolin, EJ, Howard, MJ &

979

Heuckeroth, RO 2020b, 'Robust, 3-Dimensional Visualization of Human Colon Enteric

980

Nervous System Without Tissue Sectioning', Gastroenterology, vol. 158, no. 8, pp. 2221-

981

2235.e2225.

982

983

Haase, R, Jain, A, Rigaud, S, Vorkel, D, Rajasekhar, P, Suckert, T, Lambert, TJ, Nunez-

984

Iglesias, J, Poole, DP, Tomancak, P & Myers, EW 2020, 'Interactive design of GPU-

985

accelerated Image Data Flow Graphs and cross-platform deployment using multi-lingual code

986

generation', bioRxiv, vol., p. 2020.2011.2019.386565.

987

988

Haase, R, Lambert, T, Rajasekhar, P, grahamross123, Nunez-Iglesias, J, Whitehead, L,

989

Avenel, C & Sandaltzopoulou, E 2022, py-clesperanto,

990

https://github.com/clEsperanto/pyclesperanto_prototype

991

992

Haase, R, Royer, LA, Steinbach, P, Schmidt, D, Dibrov, A, Schmidt, U, Weigert, M,

993

Maghelli, N, Tomancak, P, Jug, F & Myers, EW 2020, 'CLIJ: GPU-accelerated image

994

processing for everyone', Nature Methods, vol. 17, no. 1, pp. 5-6.

995

996

Hamnett, R, Dershowitz, L, Gomez-Frittelli, J & Kaltschmidt, J 2022, Regional

997

cytoarchitecture of the adult and developing mouse enteric nervous system, Images, zenodo.

998

999

Hamnett, R, Dershowitz, LB, Sampathkumar, V, Wang, Z, Gomez-Frittelli, J, De Andrade,

1000

V, Kasthuri, N, Druckmann, S & Kaltschmidt, JA 2022, 'Regional cytoarchitecture of the

1001

adult and developing mouse enteric nervous system', Curr Biol, vol. 32, no. 20, pp. 4483-

1002

4492 e4485.

1003

1004

Heuckeroth, RO 2018, 'Hirschsprung disease — integrating basic science and clinical

1005

medicine to improve outcomes', Nature Reviews Gastroenterology & Hepatology, vol. 15,

1006

no. 3, pp. 152-167.

1007

1008

Howard, MJ 2021, 3D imaging of enteric neurons in mouse SPARC Consortium.

1009

1010

Ingrid, D, Tatsuhiro, M, Sebastien, K, Gert De, H, Karel, G, Ricard, F, Pieter Vanden, B &

1011

Jan, T 2013, 'Gastrointestinal Motility Changes and Myenteric Plexus Alterations in

1012

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Spontaneously Diabetic Biobreeding Rats', Journal of Neurogastroenterology and Motility,

1013

vol. 19, no. 2, pp. 161-170.

1014

1015

Jacquemet, G 2021, 'Deep learning to analyse microscopy images', The Biochemist, vol. 43,

1016

no. 5, pp. 60-64.

1017

1018

Kapur, RP 2013, 'Counting neurons is not as easy as ‘one-two, three’',

1019

Neurogastroenterology & Motility, vol. 25, no. 7, pp. 549-553.

1020

1021

Kobayashi, Y, Bukowski, A, Das, S, Espenel, C, Gomez-Frittelli, J, Wagle, N, Bakshi, S,

1022

Saha, M, Kaltschmidt, JA, Venkataraman, A & Kulkarni, S 2021, 'COUNTEN, an AI-Driven

1023

Tool for Rapid and Objective Structural Analyses of the Enteric Nervous System', eNeuro,

1024

vol. 8, no. 4.

1025

1026

Li, Z, Hao, MM, Van den Haute, C, Baekelandt, V, Boesmans, W & Vanden Berghe, P 2019,

1027

'Regional complexity in enteric neuron wiring reflects diversity of motility patterns in the

1028

mouse large intestine', eLife, vol. 8.

1029

1030

Lowe, DG 2004, 'Distinctive Image Features from Scale-Invariant Keypoints', International

1031

Journal of Computer Vision, vol. 60, no. 2, pp. 91-110.

1032

1033

McQuade, RM, Singleton, LM, Wu, H, Lee, S, Constable, R, Di Natale, M, Ringuet, MT,

1034

Berger, JP, Kauhausen, J, Parish, CL, Finkelstein, DI, Furness, JB & Diwakarla, S 2021, 'The

1035

association of enteric neuropathy with gut phenotypes in acute and progressive models of

1036

Parkinson’s disease', Scientific Reports, vol. 11, no. 1, p. 7934.

1037

1038

McQuin, C, Goodman, A, Chernyshev, V, Kamentsky, L, Cimini, BA, Karhohs, KW, Doan,

1039

M, Ding, L, Rafelski, SM & Thirstrup, D 2018, 'CellProfiler 3.0: Next-generation image

1040

processing for biology', PLoS biology, vol. 16, no. 7, p. e2005970.

1041

1042

Michel, K, Kuch, B, Dengler, S, Demir, IE, Zeller, F & Schemann, M 2022a, 'How big is the

1043

little brain in the gut? Neuronal numbers in the enteric nervous system of mice, Guinea pig,

1044

and human', Neurogastroenterology & Motility, vol. 34, no. 12, p. e14440.

1045

1046

Michel, K, Kuch, B, Dengler, S, Demir, IE, Zeller, F & Schemann, M 2022b, 'How big is the

1047

little brain in the gut? Neuronal numbers in the enteric nervous system of mice, Guinea pig,

1048

and human', Neurogastroenterology & Motility, vol. n/a, no. n/a, p. e14440.

1049

1050

Nestor-Kalinoski, A, Smith-Edwards, KM, Meerschaert, K, Margiotta, JF, Rajwa, B, Davis,

1051

BM & Howard, MJ 2022, 'Unique Neural Circuit Connectivity of Mouse Proximal, Middle,

1052

and Distal Colon Defines Regional Colonic Motor Patterns', Cellular and Molecular

1053

Gastroenterology and Hepatology, vol. 13, no. 1, pp. 309-337.e303.

1054

1055

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Niesler, B, Kuerten, S, Demir, IE & Schäfer, K-H 2021, 'Disorders of the enteric nervous

1056

system — a holistic view', Nature Reviews Gastroenterology & Hepatology, vol. 18, no. 6,

1057

pp. 393-410.

1058

1059

Ollion, J, Cochennec, J, Loll, F, Escudé, C & Boudier, T 2013, 'TANGO: a generic tool for

1060

high-throughput 3D image analysis for studying nuclear organization', Bioinformatics, vol.

1061

29, no. 14, pp. 1840-1841.

1062

1063

Pachitariu, M & Stringer, C 2022, 'Cellpose 2.0: how to train your own model', Nature

1064

Methods, vol. 19, no. 12, pp. 1634-1641.

1065

1066

Palla, G, Spitzer, H, Klein, M, Fischer, D, Schaar, AC, Kuemmerle, LB, Rybakov, S, Ibarra,

1067

IL, Holmberg, O, Virshup, I, Lotfollahi, M, Richter, S & Theis, FJ 2022, 'Squidpy: a scalable

1068

framework for spatial omics analysis', Nature Methods, vol. 19, no. 2, pp. 171-178.

1069

1070

Poon, SSB, Hung, LY, Wu, Q, Parathan, P, Yalcinkaya, N, Haag, A, Luna, RA, Bornstein,

1071

JC, Savidge, TC & Foong, JPP 2022, 'Neonatal antibiotics have long term sex-dependent

1072

effects on the enteric nervous system', The Journal of Physiology, vol. 600, no. 19, pp. 4303-

1073

4323.

1074

1075

Ronneberger, O, Fischer, P & Brox, T 'U-net: Convolutional networks for biomedical image

1076

segmentation', Springer, pp. 234-241.

1077

1078

Rose, F, Rappez, L, Triana, SH, Alexandrov, T & Genovesio, A 2020, 'PySpacell: A Python

1079

Package for Spatial Analysis of Cell Images', Cytometry Part A, vol. 97, no. 3, pp. 288-295.

1080

1081

Schäppi, MG, Staiano, A, Milla, PJ, Smith, VV, Dias, JA, Heuschkel, R, Husby, S, Mearin,

1082

ML, Papadopoulou, A, Ruemmele, FM, Vandenplas, Y & Koletzko, S 2013, 'A practical

1083

guide for the diagnosis of primary enteric nervous system disorders', J Pediatr Gastroenterol

1084

Nutr, vol. 57, no. 5, pp. 677-686.

1085

1086

Schindelin, J, Arganda-Carreras, I, Frise, E, Kaynig, V, Longair, M, Pietzsch, T, Preibisch, S,

1087

Rueden, C, Saalfeld, S, Schmid, B, Tinevez, J-Y, White, DJ, Hartenstein, V, Eliceiri, K,

1088

Tomancak, P & Cardona, A 2012, 'Fiji: an open-source platform for biological-image

1089

analysis', Nature Methods, vol. 9, no. 7, pp. 676-682.

1090

1091

Schmidt, U, Weigert, M, Broaddus, C & Myers, G 'Cell detection with star-convex polygons',

1092

Springer, pp. 265-273.

1093

1094

Semar, S, Klotz, M, Letiembre, M, Van Ginneken, C, Braun, A, Jost, V, Bischof, M,

1095

Lammers, WJ, Liu, Y, Fassbender, K, Wyss-Coray, T, Kirchhoff, F & Schäfer, K-H 2013,

1096

'Changes of the Enteric Nervous System in Amyloid-

Protein Precursor Transgenic Mice

1097

Correlate with Disease Progression', Journal of Alzheimer's Disease, vol. 36, pp. 7-20.

1098

1099

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Shaw, PJ 2006, 'Comparison of Widefield/Deconvolution and Confocal Microscopy for

1100

Three-Dimensional Imaging', in JB Pawley (ed.), Handbook Of Biological Confocal

1101

Microscopy, Springer US, Boston, MA, <

https://doi.org/10.1007/978-0-387-45524-2_23

1102

pp. 453-467.

1103

1104

Sofroniew, N, Lambert, T, Evans, K, Nunez-Iglesias, J, Bokota, G, Winston, P,

1105

Peña

Castellanos, G, Yamauchi, K, Bussonnier, M & Doncila Pop, D 2022, 'napari: a multi-

1106

dimensional image viewer for Python', Zenodo, vol.

1107

1108

Stoltzfus, CR, Filipek, J, Gern, BH, Olin, BE, Leal, JM, Wu, Y, Lyons-Cohen, MR, Huang,

1109

JY, Paz-Stoltzfus, CL, Plumlee, CR, Pöschinger, T, Urdahl, KB, Perro, M & Gerner, MY

1110

2020, 'CytoMAP: A Spatial Analysis Toolbox Reveals Features of Myeloid Cell

1111

Organization in Lymphoid Tissues', Cell Reports, vol. 31, no. 3, p. 107523.

1112

1113

Stringer, C, Wang, T, Michaelos, M & Pachitariu, M 2021, 'Cellpose: a generalist algorithm

1114

for cellular segmentation', Nature Methods, vol. 18, no. 1, pp. 100-106.

1115

1116

Swaminathan, M & Kapur, RP 2010, 'Counting myenteric ganglion cells in histologic

1117

sections: an empirical approach', Hum Pathol, vol. 41, no. 8, pp. 1097-1108.

1118

1119

Tasnadi, EA, Toth, T, Kovacs, M, Diosdi, A, Pampaloni, F, Molnar, J, Piccinini, F &

1120

Horvath, P 2020, '3D-Cell-Annotator: an open-source active surface tool for single-cell

1121

segmentation in 3D microscopy images', Bioinformatics, vol. 36, no. 9, pp. 2948-2949.

1122

1123

Tseng, Q, Wang, I, Duchemin-Pelletier, E, Azioune, A, Carpi, N, Gao, J, Filhol, O, Piel, M,

1124

Théry, M & Balland, M 2011, 'A new micropatterning method of soft substrates reveals that

1125

different tumorigenic signals can promote or reduce cell contraction levels', Lab on a Chip,

1126

vol. 11, no. 13, pp. 2231-2240.

1127

1128

Vaezi, MF, Felix, VN, Penagini, R, Mauro, A, de Moura, EGH, Pu, LZCT, Martínek, J &

1129

Rieder, E 2016, 'Achalasia: from diagnosis to management', Annals of the New York Academy

1130

of Sciences, vol. 1381, no. 1, pp. 34-44.

1131

1132

Van Ginneken, C, Schäfer, K-H, Van Dam, D, Huygelen, V & De Deyn, PP 2011,

1133

'Morphological changes in the enteric nervous system of aging and APP23 transgenic mice',

1134

Brain Research, vol. 1378, pp. 43-53.

1135

1136

von Chamier, L, Laine, RF, Jukkala, J, Spahn, C, Krentzel, D, Nehme, E, Lerche, M,

1137

Hernández-Pérez, S, Mattila, PK, Karinou, E, Holden, S, Solak, AC, Krull, A, Buchholz, T-

1138

O, Jones, ML, Royer, LA, Leterrier, C, Shechtman, Y, Jug, F, Heilemann, M, Jacquemet, G

1139

& Henriques, R 2021, 'Democratising deep learning for microscopy with ZeroCostDL4Mic',

1140

Nature Communications, vol. 12, no. 1, p. 2276.

1141

1142

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint

Gut Analysis Toolbox

Wang, L, Yuan , P-Q, Gould, T & Tache, Y 2021, Antibodies tested in the colon – Mouse,

1143

SPARC Consortium.

1144

1145

Wolny, A, Cerrone, L, Vijayan, A, Tofanelli, R, Barro, AV, Louveaux, M, Wenzl, C, Strauss,

1146

S, Wilson-Sánchez, D, Lymbouridou, R, Steigleder, SS, Pape, C, Bailoni, A, Duran-Nebreda,

1147

S, Bassel, GW, Lohmann, JU, Tsiantis, M, Hamprecht, FA, Schneitz, K, Maizel, A &

1148

Kreshuk, A 2020, 'Accurate and versatile 3D segmentation of plant tissues at cellular

1149

resolution', eLife, vol. 9, p. e57613.

1150

1151

Yang, T, Ozcoban, V, Pasam, A, Kocovski, N, Pizzolla, A, Huang, Y-K, Bass, G, Keam, SP,

1152

Neeson, PJ, Sandhu, SK, Goode, DL & Trigos, AS 2020, 'SPIAT: An R package for the

1153

Spatial Image Analysis of Cells in Tissues', bioRxiv, vol., p. 2020.2005.2028.122614.

1154

1155

1156

CC-BY 4.0 International license

available under a

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint

this version posted January 20, 2024.

;

https://doi.org/10.1101/2024.01.17.576140

doi:

bioRxiv preprint