Analysis of Regional Variations of the Interstitial Cells of Cajal in the Murine Distal Stomach Informed by Confocal Imaging and Machine Learning Methods

Abstract

Introduction

The network of Interstitial Cells of Cajal (ICC) plays a plethora of key roles in maintaining, coordinating, and regulating the contractions of the gastrointestinal (GI) smooth muscles. Several GI functional motility disorders have been associated with ICC degradation. This study extended a previously reported 2D morphological analysis and applied it to 3D spatial quantification of three different types of ICC networks in the distal stomach guided by confocal imaging and machine learning methods. The characterization of the complex changes in spatial structure of the ICC network architecture contributes to our understanding of the roles that different types of ICC may play in post-prandial physiology, pathogenesis, and/or amelioration of GI dsymotility- bridging structure and function.

Methods

A validated classification method using Trainable Weka Segmentation was applied to segment the ICC from a confocal dataset of the gastric antrum of a transgenic mouse, followed by structural analysis of the segmented images.

Results

The machine learning model performance was compared to manually segmented subfields, achieving an area under the receiver-operating characteristic (AUROC) of 0.973 and 0.995 for myenteric ICC (ICC-MP; n = 6) and intramuscular ICC (ICC-IM; n = 17). The myenteric layer in the distal antrum increased in thickness (from 14.5 to 34 μm) towards the lesser curvature, whereas the thickness decreased towards the lesser curvature in the proximal antrum (17.7 to 9 μm). There was an increase in ICC-MP volume from proximal to distal antrum (406,960 ± 140,040 vs. 559,990 ± 281,000 μm³; p = 0.000145). The % of ICC volume was similar for ICC-LM and for ICC-CM between proximal (3.6 ± 2.3% vs. 3.1 ± 1.2%; p = 0.185) and distal antrum (3.2 ± 3.9% vs. 2.5 ± 2.8%; p = 0.309). The average % volume of ICC-MP was significantly higher than ICC-IM at all points throughout sample (p < 0.0001).

Conclusions

The segmentation and analysis methods provide a high-throughput framework of investigating the structural changes in extended ICC networks and their associated physiological functions in animal models.

Introduction

Interstitial Cells of Cajal (ICC) are specialized pacemaker cells distributed throughout the gastrointestinal (GI) tract to generate bioelectrical slow waves (SWs) and coordinate motility.^17,25 Multiple histopathological and confocal imaging evidences have demonstrated significant ICC changes associated with various GI motility disorders.²⁰ In particular, histopathological findings have revealed that dysmotility is strongly associated with reduced ICC numbers.¹² The assessment of ICC at present is limited by the lack of readily available quantitative metrics. A framework for quantifying and analyzing these ICC networks could provide a valuable tool to elucidate the pathophysiology of GI motility disorders.

The various types of ICC networks can be classified based on morphology, anatomical locations within the GI tissue, ultrastructural features, and function.²⁰ The ICC in the myenteric plexus (ICC-MP) are located between the longitudinal and circular muscle layers¹³ and comprise multipolar cells.⁷ Intramuscular ICC (ICC-IM) are located within the muscle layers and are associated with nerve varicosities throughout these smooth muscle layers.^13,17 The ICC-IM of the circular muscle layer (ICC-CM) are abundant bipolar cells that run circumferentially parallel with the circular muscle fibers,⁷ while the ICC-IM of the longitudinal muscle layer (ICC-LM) are sparse bipolar cells that run in parallel with the longitudinal muscle fibers.⁷ The ICC-MP are the primary pacemaker cells in the gastric antrum and small intestine,⁹ while ICC-IM provide secondary pacemaker activity of slow waves, and supports circumferential SW propagation.¹⁶ Experimental recordings and modeling evidence have demonstrated that ICC-MP also serve as a bidirectional “coupler” of ICC-LM and ICC-CM in order to achieve a uniform antegrade propagation around the stomach.¹⁰

In the present study, a machine learning method was employed to automatically segment ICC-MP, and ICC-LM and ICC-CM in the antrum of a transgenic mouse. Structural metrics were then applied to assess structural variations of ICC populations around the circumference, i.e., from greater to lesser curvature, and along the length of the gastric antrum. The analysis extended on a previous preliminary study reported in a conference proceeding that applied a number of classifiers in a small image field,²¹ to accurately segment three different ICC networks in larger imaging fields with additional metrics for validation and quantification of spatial variations. The study provides an efficient and accurate framework for the analysis of ICC of the gut, with specific applications in addressing the complex relationship between variations of ICC structures and distinct functions of different portions of the stomach.

Methods

Whole-Mount ICC Image Acquisition

All animal studies were carried out in accordance with international guidelines and approved by the Université Libre de Bruxelles Animal Welfare Committee (Protocols 491N and 552N). A whole-mount tissue sample was obtained from the stomach antrum of a 10 week old male (c-KitCreERT2, R26mT-mG) transgenic mouse that expressed enhanced green fluorescent protein in the c-Kit positive cells after administration of tamoxifen.¹⁹ C-Kit is a receptor tyrosine kinase, which is specifically expressed in ICC within the gut musculature and serves as a fiducial marker for ICC.³⁴

The mouse was orally administered 5 mg of tamoxifen for 4 days and the gastric antrum was collected 48 hours after the last administration. Then, the specimen was dissected along the lesser curvature, pinned flat on SYLGARD™ Silicone Elastomer (Dow Chemical Company, Zwijndrecht, Belgium) in a Petri dish, and fixed overnight in fresh 4% paraformaldehyde pH 7.4 and stored in 0.01M PBS/azide 0.1% (w/v). Finally, CUBIC tissue clearing was performed as detailed in a previous study,³⁰ and the sample was equilibrated in refractive index matching solution (RIMS, Histodenz, Sigma-Aldrich, cat. no. D2158).³³

Two groups of transverse image tiles were acquired using a multiphoton microscope (LSM780NLO, Zeiss, Iena, Germany) with an excitation wavelength of 920 nm, LD C-Apochromat 40×/1.1 water immersion objective, and a digital zoom of 0.7. Frames of 512 × 512 pixels with a field of view of 303 × 303 μm² were acquired as an image stack with 1 μm spacing across the thickness of gastric wall. The proximal (6170 × 862 × 285 μm³) and distal (5611 × 862 × 266 μm³) tiles were composed of a grid of image stacks with 8% overlap and their approximate locations are illustrated in Fig. 1a. A total of 126 whole-mount image stacks were analyzed in this study (n = 66 in proximal antrum, and n = 60 in distal antrum).

Figure 1.

The ICC image segmentation workflow for extracting, training, validating the FRF model. (a) The antrum was dissected along the lesser curvature and the approximate location of two transverse tissue strips (proximal and distal) within the mouse gastric antrum is shown. (b) Training and validation of FRF Model: An example of a set of selected training sub-images (outlined in blue) from the 11 selected image stacks (I to V) to generate one FRF model. (c) Flow chart depicting the steps taken in performing the FRF segmentation on the ICC networks.

Image Segmentation

A previously described machine learning approach was used to segment ICC from each image stack.²¹ A “gold standard” (GS) training dataset was created by manually segmenting ICC from non-ICC regions from 17 sub-images of 59.3 × 59.3 μm² (100 × 100 pixels). Three investigators performed the manual segmentation independently and the median of the three binary ICC masks was defined as the overall GS training dataset.

A Fast Random Forest (FRF) classification method implemented as part of the Trainable Weka Segmentation tool in ImageJ was used to segment the ICC structures.¹ A total of 17 GS images were utilized, from five arbitrary selected confocal image stacks obtained from across the proximal antrum (Fig. 1a). For each of these arbitrary selected confocal image stacks (I to V), one image slice (512 × 512 pixels) from the MP region was randomly selected and cropped into smaller sub-images of 100 × 100 pixels, i.e., 3 sub-images each from image stacks I–IV; and 5 sub-images from image stack V as illustrated in Fig. 1b. In each iteration, a total of 11 images which comprised of two images randomly selected from image stacks I–IV, and three randomly selected from image stack V were used for training, with the remaining 6 used as validation. The process was repeated for all 810 combinations of different training and validation sets. For each iteration, the performance of the FRF model was assessed using the following metrics:

$$\mathrm{Dice}\mathrm{Similarity}\mathrm{Coefficient}\left(\mathrm{DSC}\right)=\frac{2\left|\mathrm{GS},\cap ,I\right|}{\left|\mathrm{GS}\right|+\left|I\right|}$$ 1

$$\mathrm{Jaccard}\mathrm{Index}\left(\mathrm{JI}\right)=\frac{\left|G,S,\cap ,I\right|}{\left|G,S,\cup ,I\right|}$$ 2

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$ 3

$$\mathrm{Rand}\mathrm{Index}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$$ 4

$$\mathrm{Sensitivity}\left(\mathrm{Recall}\right)=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$ 5

$$\mathrm{Specificity}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$$ 6

$$\mathrm{Accuracy}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}$$ 7

$$F-\mathrm{Measure}=2\times \frac{\mathrm{Recall}\times \mathrm{Precision}}{\mathrm{Recall}+\mathrm{Precision}}$$ 8

where |GS| is the number of foreground pixels in the gold standard image, |I| is the number of foreground pixels detected as ICC in the binary mask resulting from the FRF classification model. TP, TN, FP, FN correspond to true positive, true negative, false positive, false negative, respectively. The top performing model was chosen based on the ranking of the sums of the eight metrics and used to segment all 126 confocal image stacks.

ICC Network Analysis

Examining the confocal image stacks from the serosal to the mucosal surface: the ICC-LM is closest to the serosal surface, followed by the ICC-MP and ICC-CM, as shown in Fig. 3b. The thickness of each tissue layer along the gastric wall was manually demarcated for each image stack. Metrics including density, alignment index (AI), and orientation were applied to quantify the ICC structures. The density is defined as the ratio of the number of pixels representing ICC over the total number of pixels in the image, expressed as a percentage. The AI metric represents the overall relative alignment of the ICC processes in the image plane, with “0” indicating all the ICC processes are not aligned with one another, and “1” indicating all the ICC processes are perfectly aligned in the image. The orientation metric measured the predominant orientation of the ICC processes measured from the transverse axis, which ranges from 0° to 180°. In addition, the percentage volume is defined as the ratio of the number of ICC voxels over the total number of voxels in the confocal image stack.

Figure 3.

2D ICC network morphology and quantitative analysis following segmentation of ICC. (a) An isometric view of a single image stack from the transverse distal strip near the greater curvature. (i) Z-projection of the segmented image stack based on the standard deviation of pixels along the circumferential-transmural plane and (ii) along the longitudinal-transmural plane. The ICC-LM, ICC-MP, and ICC-CM can be observed within the longitudinal muscle (LM), myenteric plexus (MP), and circumferential muscle (CM) layers, respectively. (b) Structural quantification of ICC network showing variation in % density, alignment index, and orientation of ICC from serosa to mucosa

Alignment Index (AI) was determined by performing a 2D discrete Fourier transform of the segmented image, with the zero-frequency component, i.e., Direct Coefficient, shifted to the center of the matrix. Thereafter, the power spectrum was computed and subsequently converted to polar coordinates of 1° intervals. The spectrum was bandpass filtered between 5 and 128 intensity values to eliminate center pixels and to avoid aliasing. For each 1° intervals, the averaged line intensities were computed, I(θ), i.e, the average of all pixels forming a particular angle with respect to the horizontal axis, based on Eq. (9).

$$I\left(\theta \right)=\frac{{\int }_{\theta }^{\theta +\mathrm{\Delta }\theta }{\int }_{r=5}^{128}{\mathrm{PS}}_{r,\theta }\mathrm{d}r\mathrm{d}\theta }{N_{\mathrm{Total}}}$$ 9

where I is the average line intensities for angles 0°<θ < 360°, with Δθ = 1°; PS_r,θ is the power spectrum given in polar coordinates and N_total is the total number of pixels in the image.

The averaged line intensity, I(θ) was transformed to cartesian coordinates (Eq. (10)) and subsequently, the AI was calculated from eigenvalues of the normalized covariance matrix, C, as follows:

$${\mathit{C}}_{\mathit{x}\mathit{y}}=[I\left(\theta \right)\cos \left(\theta \right);I\left(\theta \right)\sin \left(\theta \right)];\mathrm{for}{0}^{\circ }<\theta <{360}^{\circ }$$ 10

$$\mathit{C}={\mathit{C}}_{\mathit{x}\mathit{y}}^{{}^{\mathbf{\prime }}}\times {\mathit{C}}_{\mathit{x}\mathit{y}}$$ 11

$$AI=1-\frac{{\lambda }_2}{{\lambda }_1}$$ 12

where λ₁ and λ₂ are the eigenvalues of C.

The eigenvalues and eigenvector of the C for the power spectrum were used to calculate the orientation. The unit eigenvector V, was projected onto the mapping matrix such that the predominant orientation of the ICC networks on the image can be given as:

$$\theta ={\tan }^{-1}(\frac{V_{11}}{V_{12}})$$ 13

where V₁₁ and V₁₂ are elements of V.

The AI, density, and orientation metrics were evaluated for the ICC-LM, ICC-MP, and ICC-CM networks classified in the 126 sampled image stacks.

Fractal analysis was performed based on Hausdorff–Besicovitch dimension by the box-counting method,^22,37 which involved partitioning the FRF segmented binary image of the ICC networks into several grid of boxes of equal size ε × ε. The procedure was performed for different values of ε ranging from 2 to 512 (image size) in increments of powers of 2. Boxes containing at least one ICC pixel were considered positive, and those without any ICC pixels were considered negative. The number of positive boxes for a given box width ε is given as:

$$N\left(\epsilon \right)\approx {\epsilon }^{-D_0}$$ 14

where D₀ denotes the Hausdorff–Besicovitch dimension, which is defined as the slope of the linear region in the log–log plot of N(ε) versus ε.

The Hausdorff–Besicovitch dimension describes how much of the image is filled by ICC, while the lacunarity measure describes the spatial patterns of the imaged ICC network.²⁸ Lacunarity was measured using a gliding box algorithm. The gliding box algorithm was applied by shifting an ε × ε box starting at the top left corner of an image and counting the number of ICC pixels within the box. The process is repeated as the box is shifted by one pixel to the right each time until reaching the bottom right corner of the image.³² The lacunarity, L(ε) for each box width ε is calculated as:

$$L\left(\epsilon \right)=\frac{N\left(\epsilon \right)\times Q_2}{Q_1^2}$$ 15

$$Q_1={\sum }_{i}p(i,\epsilon )$$ 16

$$Q_2={\sum }_{i}p(i,\epsilon {)}^2$$ 17

where N(ε) is the number of boxes of width ε given the size of the image, x as N(ε) = (x−ε + 1)², Q₁ is the sum of the number of signal pixels in each box, Q₂ is the sum of the square of the number of signal pixels in each box, and p(i,ε) is the number of signal pixels on the ith box and i ∈[1, N(ε)]. Each 2D image was assigned 9 values of ε and line of best fit was determined from L(ε) against ε when plotted on a natural log–log scale.

$$\ln \left(L,\left(ϵ\right)\right)=-D_0\ln (\epsilon )+\ln \left(\frac{Q_2}{Q_1^2}\right)$$ 18

Statistical Analysis

Statistical analysis was performed in MATLAB (R2019b, MathWorks, Natick, MA, USA). Unless specified, data are reported in mean ± standard deviation. A one-way ANOVA was applied to each quantitative metric to determine significant difference between ICC-LM, ICC-MP, and ICC-CM within the proximal and distal antrum regions, respectively. Where significance was found, a post-hoc test with Bonferroni correction was performed to reveal pairwise differences. A two-sample t-test or Mann–Whitney U test was performed to determine the specific relationship of statistical significance further between proximal and distal ICC. The adjusted p < 0.05 was considered statistically significant. The “n” values reported in the text refer to the number of confocal tissue stacks that makes up the proximal/distal antral tissue strip. The proximal antral region was made of 66 stacks of images with ICC-LM (n = 32), ICC-MP (n = 66), and ICC-CM (n = 66). The distal antral region was made of 60 stacks of images, in which 5 confocal distal stacks were excluded from the analysis due to poor image quality resulting in ICC-LM (n = 38), ICC-MP (n = 55), and ICC-CM (n = 55).

Results

The variations in performance amongst the 810 FRF models were minimal. The average performance of the 810 FRF models was evaluated for its Dice Coefficient (94.15 ± 0.44 %), Jaccard Index (89.0 ± 0.8 %), Rand Index (93.5 ± 0.5 %), sensitivity (94.3 ± 1.0 %), specificity (91.7 ± 1.8 %), precision (94.2 ± 1.0 %), accuracy (89.0 ± 0.8 %) and F-measure (94.2 ± 0.4 %). As portrayed in Fig. 2b, only minor differences were observed between the best and the worst performing FRF models, which demonstrates the robustness of the FRF model in performing the ICC segmentation. The best performing FRF model amongst the 810 had a Dice Coefficient of 95.4%, Jaccard Index of 91.1%, Rand Index of 94.7%, sensitivity of 95.6%, specificity of 93.0%, precision of 95.2%, accuracy of 91.1%, and F-measure of 95.4%. This model was used to segment the ICC networks for the entire transverse proximal and distal confocal tissue strips across the imaged mouse antrum.

Figure 2.

Performance of the FRF model. (a) Histogram of the performance metrics evaluated for the 810 FRF models. (b) Spider chart depicting evaluation for the best and worst performing models. (c) The receiver operating characteristic (ROC) curve for the best performing FRF model segmentation of ICC-MP, n = 6 (AUROC = 0.9731) and ICC-IM, n = 17 (AUROC = 0.9954) networks.

To quantify the performance of the proposed Weka FRF model, the area under the ROC curve (AUROC) was calculated, as shown in Fig. 2c. While the ICC-IM was not used as a training dataset, the FRF model nevertheless produced comparable performance to the ICC-MP segmentation (AUROC: 0.995 vs 0.973). The high AUROC in both ICC types demonstrated the robustness and excellent performance of the FRF model to segment the entire tissue stacks using the same classifier.

An example of the segmented image stack is shown in Fig. 3a, where the ICC-LM, ICC-MP and ICC-CM layers are visible. When transmural cross-sectional views of along the circumferential direction (Fig. 3 a-i) and longitudinal direction (Fig. 3 a-ii) are viewed, it is clear that ICC-MP formed the densest layer, from which protrusions of septal ICC were also visible (Fig. 3a-ii). The % density of ICC was also measured in the transmural direction and the ICC-MP layer contained significantly denser ICC than the surrounding layers, shown in Fig. 3b. In addition, due to the bidirectional nature of the ICC-MP layer, the alignment index of ICC falls as it approaches the ICC-MP layer, whereas it is closer to “1” in the ICC-LM and ICC-CM layers. There was also an approximately 90° shift in the orientation of the ICC between ICC-LM and ICC-CM network.

Lacunarity analysis was applied to each identified image slice of ICC-LM, ICC-MP and ICC-CM networks from a confocal stack. The average lacunarity measure over all cohort sets of images of ICC-LM, ICC-MP and ICC-CM networks for each stack was obtain. Subsequently, the mean lacunarity, L(ε) for all 66 proximal and 60 distal image stacks respectively was plotted over a range of given box sizes, ε from 2 to 512. The lacunarity plots shown in Fig. 4 is given as log–log plots of L(ε) versus ε for ICC-LM, ICC-MP and ICC-CM with error bars representing the standard error over the n = 66 proximal and n = 60 distal cohort sets of image stacks. The lacunarity measure serves as another insightful metric for the quantitative analysis of ICC network. With reference to Eq. (18), the slope of the plots in Fig. 4 corresponds to the Hausdorff–Besicovitch dimension (D₀) and can be used to identify the extent of irregularity in asymmetrical and heterogeneous ICC network images across the proximal and distal antrum regions. The slope describes how much space is filled, while the lacunarity describes the extent of distribution of the sizes of gaps surrounding the ICC within the imaged network. A high lacunarity describes an ICC network with more prevalent distribution of gaps throughout.

Figure 4.

Lacunarity Analysis. Log–log plots for the average lacunarity of all identified ICC-LM, ICC-MP & ICC-CM image slices in all 66 proximal and 60 distal image stacks. The proximal and distal lacunarity plots overlaps at higher dimensions of log (ε) for all three classes of ICC networks.

The fractal dimension of ICC-LM was found to be slightly lower in the proximal compared to the distal antrum (1.9 × 10⁻³ vs. 2.2 × 10⁻³). On the other hand, the slope for ICC-MP and ICC-CM in the proximal region was greater compared to the distal region; 4.8 × 10⁻² vs. 3.8 × 10⁻² for ICC-MP and 2.8 × 10⁻³ vs. 2.0 × 10⁻³ for ICC-CM networks. Unlike the sparse, striated strands of ICC-LM and ICC-CM networks, the mean lacunarity measure for the “web-like” ICC-MP networks was significantly higher (p < 0.0001) in both proximal and distal antral regions. Thus, the results revealed a significantly higher distribution of gaps and a difference in the distribution of ICC within the MP layer.

With reference to the box plots in Fig. 5, the densities of ICC-LM and ICC-CM were significantly lower than ICC-MP in both proximal and distal antral tissues (both p < 0.0001). On the other hand, the alignment indexes of ICC-LM and ICC-CM were significantly higher from ICC-MP with both reporting p < 0.0001. The relative orientations of ICC-MP was significantly different from the relative orientations of ICC-LM, in the proximal (p < 0.0001) and distal (p = 0.00188) antrum regions respectively.

Figure 5.

Statistical analysis performed for ICC-LM, ICC-MP, and ICC-CM in the proximal and distal antral regions, respectively by one-way ANOVA for each quantitative metric. The relative orientation is defined as the orientation of ICC-LM/ICC-MP relative to ICC-CM network processes. A statistically significant difference was found in all 6 sub-figures (** p< 0.0001, *p < 0.001). A plus symbol in the boxplot represents an outlier.

The results of the two-sample t-test are summarized in Table 1. The % density of ICC-MP and ICC-CM between the proximal and distal antrum regions were significantly different with p = 0.0223 and p < 0.0001, respectively. The alignment indexes of ICC-LM and ICC-CM between the proximal and distal antrum regions were significantly different with p = 0.00790 and p < 0.0001, respectively. The orientation of ICC-LM relative to ICC-CM network processes were significantly different between the proximal and distal antrum regions with p = 0.0433. On the other hand, no significant difference was found between the proximal and distal antrum regions for the % density of ICC-LM (p = 0.401), the AI of ICC-MP (p = 0.948) and the relative orientation of ICC-MP (p = 0.972).

Table 1.

Statistical summary table for two-sample t-test.

Groups	ICC-LM		ICC-MP		ICC-CM
Regions	Proximal	Distal	Proximal	Distal	Proximal	Distal
Density (%)	1.24 ± 0.76	1.46 ± 1.30	23.10 ± 10.08	18.53 ± 11.58	1.83 ± 0.44	1.26 ± 0.63
	p = 0.401		p = 0.0223		p = 1.76 × 10−7
Alignment Index	0.93 ± 0.05	0.96 ± 0.04	0.73 ± 0.13	0.73 ± 0.14	0.96 ± 0.02	0.93 ± 0.04
	p = 0.00790		p = 0.948		p = 5.10 × 10−5
Relative Orientation (°)	92.43 ± 18.17	83.98 ± 14.54	69.11 ± 23.09	68.96 ± 24.50	N/A
	p = 0.0433		p = 0.972

Figure 6a depicts the variation in the imaged intramuscular ICC (ICC-LM and ICC-CM) network orientation across the transverse proximal and distal antrum regions. The networks of ICC-LM and ICC-CM run parallel to the longitudinal and circular muscle fibers respectively.^5,8 At 2665 μm (proximal) and 2386 μm (distal), ICC-LM was oriented in the perpendicular direction to the ICC-CM, which likely indicates the location of the greater curvature where the preparation experienced least deformation. This location is demarcated by a red marker as shown in Figs. 6b and 6c.

Figure 6.

ICC spatial distribution in the mouse antrum. (a) Orientation of ICC-LM and ICC-CM across the mouse antrum as observed in the (i) proximal and (ii) distal image stacks. (b) Percentage ICC network volume distribution in the proximal and distal antral regions for: (i) ICC-LM, (ii) ICC-MP, and (iii) ICC-CM. (c) The thickness of the (i) LM layer (ii) MP layer, and (iii) CM layer in the proximal and distal antral regions.

Figure 6b shows the variation in the % of ICC network volume within the (i) LM layer, (ii) MP layer, and (iii) CM layer in the proximal and distal antrum. The mean % of ICC-LM volume was observed to be 3.6 ± 2.7 % in the proximal and 3.2 ± 3.9 % in the distal antrum. The mean % of ICC-CM volume was within similar range to % of ICC-LM volume observed with a value of 3.1 ± 1.2 % in the proximal and 2.5 ± 2.8% in the distal antrum. In the proximal antrum, the overall % volume of ICC-MP was significantly higher than the % volume of ICC-LM (30.1 ± 12.1% vs. 3.2 ± 3.9%; p < 0.0001) and ICC-CM (30.1 ± 12.1% vs. 3.1 ± 1.2%; p < 0.0001). Similarly, in the distal antrum, the overall % volume of ICC-MP was significantly higher than the % volume of ICC-LM (32.4 ± 15.7 vs. 3.2 ± 3.9%; p < 0.0001) and ICC-CM (32.4 ± 15.7 vs. 2.5 ± 2.8%; p < 0.0001). There was an observable decline in % volume of ICC-MP proceeding from the greater curvature towards the lesser curvature of the anterior antrum (Fig. 6b-ii). The decrease in % volume of ICC-MP was more distinct in the distal antrum than in the proximal antrum (average of 46.3% to 17.1% vs. 29.9% to 23.7%) over a span of 2234.5 μm.

In the mouse stomach antrum investigated, the LM layer was discerned as a thin layer of approximately 3.3 ± 2.5 μm in the proximal and 11.6 ± 7.7 μm in the distal antrum. There was a noticeable increase in the overall thickness of the MP layer proceeding from the proximal to the distal antrum, whereby, the approximate thickness of the MP layer was observed to be 15.3 ± 3.8 μm in the proximal and 22.1 ± 10.6 μm thick in the distal antrum. The CM layer predominantly was the thickest amongst the 3 transmural gastric layers; with an average thickness of 87.0 ± 20.1 μm in the proximal and 69.3 ± 26.8 μm thick in the distal antrum.

Our analysis of the sampled mouse antrum tissue in the distal region revealed a pronounced increased in the thickness of the MP layer, from an average thickness of 14.5 to 34.0 μm over approximately 2235 μm originating from the greater curvature towards the lesser curvature of the anterior antrum, as shown in Fig. 6c-ii. In contrast, the proximal region of the anterior antrum revealed a slight decrease in the thickness of the MP layer, from an average thickness of 17.7 to 9.0 μm over a similar distance proceeding towards the lesser curvature.

An increase followed by a decrease in the CM layer thickness was observed emerging from the greater curvature to the lesser curvature of the anterior antrum, in both the proximal and distal antrum regions (Fig. 6c-iii). In particular, the CM layer in the distal region, showed a gradual increase in thickness from 36.0 to 102.7 μm over a span of 1955 μm, and subsequently decrease to 70.3 μm as it draws towards the lesser curvature. Likewise, the CM layer in the proximal region varied parabolically along the span of the anterior antrum, in which it increased from 64.7 to 104.7 μm thick over a span of 111  μm and then decreased to a thickness of 63.0 μm over a similar distance towards the lesser curvature.

Discussion

Analyses of ICC network in the gastric tissue have been largely limited to small-scale 2D images and rely on nuclei counts.²⁰ We presented an approach using the Weka FRF classification to segment networks of ICC and followed by structural analysis of the networks. The reported AUROC of 0.973 for the segmentation of ICC-MP networks, and an AUROC of 0.995 for the segmentation of ICC-IM networks demonstrates accurate and robust performance.

A dense and homogeneous population of intramuscular ICC was found in the CM region running in parallel to the circular muscle cells, but very few intramuscular ICC were found in the longitudinal muscle layer with marked heterogeneity in the distribution of these cells throughout the sampled tissue. The findings of our study provides evidence of the limited amount with marked heterogeneity in the density distribution of ICC-LM in the murine antrum (Fig. 6b-i) as reported, whereas ICC-CM is a dense and homogeneous population.¹⁴ The ICC-LM of the stomach are present in the fundus and corpus^5,29 with highest density along the greater curvature in the fundus and corpus²⁹ but are very few to almost none near the lesser curvature of the corpus and in the longitudinal muscle layer of the murine antrum.^8,29 It has been demonstrated that enteric motor neurons formed functional innervation with ICC-IM that evokes neural responses in gastric circular muscles of the fundus,^3,6,36 antrum,^4,15,31 and longitudinal muscles of the corpus.²⁹ Furthermore, Song et al. also reported reduced neural regulation near the lesser curvature and in the antrum due to the reduced ICC-LM density suggesting ICC-LM play a critical role in mediating neurotransmission.²⁹ The regional variation in ICC-LM distribution and density in the fundus, corpus, and antral regions of the murine stomach in concert with regional variation in cholinergic and nitrergic neural responses²⁹ could have contributed to differences in contractile patterns that orchestrate propulsion-retropulsion of chyme that aids mixing of gastric contents in the gastric antrum.

Our findings demonstrate substantially more ICC-CM were present compared to that of ICC-LM in the mouse proximal (231,720 ± 45,497 μm³ vs. 7964 ± 2866 μm³) and distal (108,000 ± 48,740 μm³ vs. 20,082 ± 16,169 μm³) antrum. The CM layer was approximately 19–33 times thicker in the proximal antrum and approximately 4–8 times thicker in the distal antrum relative to the LM layer. It has been demonstrated that the circumferential spread of SW is sustained by ICC-CM alone.¹⁶ In addition, ICC-CM also play a critical role in maintaining the consistent amplitude of SW throughout the gastric antrum,¹⁴ with preferential conduction in the circumferential direction observed in both animal and human studies.^16,24 This anisotropic gastric SW conduction may be attributed to the thicker CM layer that results in the formation of rings of contraction, thereby facilitating the propulsion of gastric contents towards the pylorus.

The present study also indicated that the % ICC-MP volume is relatively high at the anterior side near the greater curvature region and declines as it approaches the lesser curvature, with a more marked decrease seen in the distal antrum compared to the proximal antrum. In the mouse and guinea-pig antrum, the density of ICC-MP has been reported the highest near the greater curvature and falls markedly towards the lesser curvature.^14,23 Networks of ICC-LM and ICC-MP support the longitudinal propagation of SW. In the antrum in particular, because of the limited amount of ICC-LM, longitudinal SW conduction is mainly dependent on ICC-MP. The marked decline in the density of ICC-MP is consistent with decrease in longitudinal SW propagation velocity from the greater curvature to the lesser curvature.¹⁴ This heterogeneity in longitudinal SW propagation velocity may serve to maintain the symmetry in the circumferential activation of slow waves along the stomach.

There are a number of limitations that need to be addressed in future studies. Obtaining absolute values as a benchmark for comparison in different states of health and disease will remain challenging. The scattering nature of the gut wall prevents deep imaging and present tissue-clearing protocols induce either some swelling or shrinking of the tissue that is difficult to rigorously quantify. As such, intravital imaging of the ICC networks over time and under different pathological states in the living gut would be most ideal. More advanced techniques, such as orthogonal decomposition could be applied in large cohorts to quantify the remodeling process of ICC during aging and diseases.^35,38 Once the structural remodeling process and their implications on SW propagations are better elucidated, it will be possible to investigate the potential use of inverse modeling to pinpoint certain regions of the GI tract that may contain significant alterations to the underlying ICC structure, thus providing a more specific target for intervention.^2,18 While the performance of the FRF was excellent in the samples tested, a wider range of training data from slightly different imaging conditions and more subjects would help to improve the robustness of the classifiers, as well as offering a more comprehensive description of the spatial variations of the different ICC networks. The present 2D morphological analysis of the ICC network has proven to be able to adequately report important basic configurations of the 3 different types of ICC networks. However, to further delineate ICC-MP networks from ICC-IM networks and their interfaces requires the development of more sophisticated 3D metrics. We will employ 3D morphological analysis and quantification to overcome such issues and provide more explicit and accurate measures to capture the interconnectedness between the different types of ICC in impending future studies.

Density, AI and orientation measures of 2D network morphology were employed to study the transmural variation in ICC in whole-mount gastric tissue from serosa to mucosa, while the % ICC volume and ICC layer thickness were employed to investigate the spatial variation of ICC along the transverse and aboral direction of the gastric antrum. In addition, lacunarity measure serves as another useful metric suitable for quantifying changes in ICC networks due to remodeling,²⁶ and during postnatal development.¹¹ These automated analysis allowed for detailed quantification and unbiased comparison of ICC network morphology across a spectrum of ICC degradation that will be important for simulating ICC electrophysiology and predicting GI function. To enable a reduction in the complexity of computer intensive simulation of SW propagation over large-scale realistic virtual ICC network, the proper orthogonal decomposition of the automata model can be trialed.²⁷

Conflict of interest

Sue Ann Mah, Peng Du, Recep Avci, Jean-Marie Vanderwinden, and Leo K. Cheng have declared that no conflict of interest exists.