Segmentation of Muscularis Propria in Colon Histopathology Images Using Vision Transformers for Hirschsprung's Disease

Adrian D. C. Chan; Anthony Fuller; Dina El Demellawy; Saleh Abou-Alwan; Youssef Megahed

REVIEW 4 major objections 6 minor 15 references

Segmentation of Muscularis Propria in Colon Histopathology Images Using Vision Transformers for Hirschsprung's Disease

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A Vision Transformer achieves perfect plexus inclusion when segmenting colon muscle in Hirschsprung's disease.

desk verdict Plausible incremental ViT segmentation result, but the superiority claim rests on unmatched baselines and a test-set-tuned threshold. read the letter →

arxiv 2412.20571 v1 pith:HWJAXWHE submitted 2024-12-29 eess.IV cs.AIcs.CV

Youssef Megahed , Anthony Fuller , Saleh Abou-Alwan , Dina El Demellawy , Adrian D. C. Chan This is my paper · ORCID

classification eess.IVcs.AIcs.CV

keywords VisionTransformerHirschsprung'sdiseasemuscularispropriasegmentationhistopathologydeeplearningself-attentionPlexusInclusionRatewholeslideimages

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a Vision Transformer (ViT) can segment the muscularis propria layer of the colon in calretinin-stained histopathology images more accurately than the convolutional neural network (CNN) and k-means baselines used in earlier work on Hirschsprung's disease. On the same 30 whole-slide images from 26 patients, the ViT reports a DICE score of 89.9% and a Plexus Inclusion Rate (PIR) of 100%, meaning no myenteric plexus region is left outside the segmented muscle layer. The CNN reports 89.2% DICE and 96.0% PIR, while k-means reports 70.7% DICE and 77.4% PIR (the abstract prints 80.7% for k-means DICE). Because missing a plexus region would cause ganglion cells to be overlooked in later analysis stages, the perfect PIR is the clinically meaningful result: it shows the segmentation can support automated Hirschsprung's disease assessment. If true, this would make quantitative ganglion-cell analysis feasible without adding time burden to pathologists.

What carries the argument

The central object is the Vision Transformer itself: a model that splits each 224x224 tile into non-overlapping 16x16 patches, embeds each patch into a 768-dimensional representation, and processes the sequence through transformer encoder layers whose self-attention captures long-range relationships across the whole image, unlike the local receptive fields of CNNs. The ViT is initialized with weights from masked-autoencoder pretraining on ImageNet-1k (reference [12]) and fine-tuned with AdamW, a cosine learning-rate schedule with warmup, and data augmentation. The second mechanism is the confidence-threshold rule applied to the SoftMax output: lowering the threshold admits more pixels as muscularis, which raises PIR (fewer plexus regions excluded) at the cost of precision and DICE. The paper selects the threshold that maximizes PIR while keeping DICE acceptably high, and reports the trade-off curve in Figure 2.

What would settle it

Segment the same 30 whole-slide images with the ViT, CNN, and k-means under identical 5-fold splits, preprocessing, and confidence-threshold selection; if the ViT's DICE or PIR advantage disappears or reverses under matched conditions, the central claim fails. Alternatively, evaluate the ViT on an independent set of calretinin-stained colon WSIs with fresh ground-truth annotations and check whether PIR remains 100% and DICE remains above the CNN baseline.

Watch

Extended reading notes

Core claim

The central claim is that a pre-trained and fine-tuned Vision Transformer outperforms prior CNN and shallow-learning methods for muscularis propria segmentation in Hirschsprung's disease histopathology. The ViT, built from 16x16 pixel patches with self-attention and a linear segmentation head, was pre-trained on ImageNet-1k by masked autoencoding and then fine-tuned on 24,000 tiles from 30 whole-slide images under 5-fold cross-validation. At a confidence threshold of 0.01 the model achieves a DICE of 89.9%, precision 82.4%, recall 99.7%, and PIR 100%; lowering the threshold trades precision for completeness, and at threshold 0.4 DICE reaches 95.4% while PIR falls. The paper argues that the 100% PIR at 89.9% DICE is the right operating point because excluding any plexus region would hide ganglion cells from downstream processing. The conclusion is that ViTs are a promising tool for the segmentation stage of automated Hirschsprung's disease analysis.

Load-bearing premise

The comparison to the CNN and k-means baselines assumes those earlier results were produced under the same evaluation conditions (same cross-validation split, threshold, and preprocessing) as the ViT, but the paper does not verify that the baseline protocols match.

Editorial extensions

If this is right

The ViT can serve as the first stage of an automated Hirschsprung's disease pipeline, feeding plexus segmentation and ganglion-cell detection.
A 100% PIR means downstream stages never lose a plexus region because of muscularis segmentation, so ganglion-cell counting can proceed over the full set of plexus regions.
Pathologists could use the automated muscularis mask to focus manual review, reducing inter- and intra-rater variability in HD diagnosis.
The threshold trade-off gives a principled way to prioritize sensitivity (PIR) over precision depending on the clinical goal.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the baseline numbers come from earlier studies rather than from re-running the CNN and k-means inside the ViT's training loop, part of the reported gain could reflect protocol differences; a matched re-implementation would settle this.
The same fine-tuned ViT architecture could be adapted to the next pipeline stage (plexus segmentation) with minimal changes, since it already learns to localize tissue layers with long-range context.
If the 100% PIR result replicates across more diverse staining and scanning conditions, it would justify a clinical study correlating automated ganglion counts with surgical outcomes.
A two-pass inference strategy—high threshold for a precise muscularis mask, low threshold to catch all plexus regions—could combine high DICE and perfect PIR.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: the paper reports a plausible first ViT segmentation of muscularis propria in Hirschsprung's disease, but the central claim that the ViT surpasses the CNN is not established. The CNN and k-means numbers are lifted from prior work without any check that the evaluation protocol matched, and the ViT's 'best' operating point is chosen by sweeping the threshold on the test set. Both are fixable, and the paper is otherwise a solid incremental application.

What is genuinely useful: it is the first ViT result on this specific task and dataset, and the authors are transparent about the PIR/DICE tradeoff and the rough ground truth. The 5-fold CV over 30 WSIs is a reasonable evaluation design for a small study, and Figure 2 makes the threshold behavior clear. If the only claim were 'a ViT can segment muscularis at DICE ~90 and PIR near 100', that would be credible.

Soft spots, in order of importance. First, baseline comparability. The paper says the dataset is the same as [6], but the CNN and k-means metrics are quoted from references. There is no evidence that the 5-fold split, tile size and stride, Macenko normalization, 10x downsampling, confidence threshold, or per-WSI aggregation match the ViT pipeline. A 0.7 pp DICE gap can easily come from protocol differences. Second, the threshold curve in Figure 2 is inspected to find a threshold that gives PIR=100, and then that operating point is reported as the result. That is test-set tuning and makes the reported DICE/precision/recall optimistic. A validation-based threshold choice or a fixed threshold would make the comparison clean. Third, there are no error bars, confidence intervals, or paired significance tests. With 30 WSIs from 26 patients and an unknown number of plexus regions, 89.9 vs 89.2 could be noise. Fourth, the abstract says k-means DICE is 80.7, while Table 1 and the introduction say 70.7. Minor, but it undermines confidence in the numbers.

The central idea is not circular and the paper does not overclaim badly except for the word 'surpassing.' The right fix is a matched-protocol re-evaluation of the baselines and a properly chosen threshold with uncertainty quantified. If the authors do that, the paper would be a fair incremental contribution to computational pathology for Hirschsprung's disease.

Recommendation: send it to peer review, but with a referee who will require the re-evaluation. It is not a desk reject; it is a paper whose comparison needs real work.

Referee Report

4 major / 6 minor

Summary. This paper applies a Vision Transformer (ViT) to segment the muscularis propria in calretinin-stained colon histopathology whole-slide images for Hirschsprung's disease. The authors use the same 30-WSI dataset from prior work [5,6], apply Macenko color normalization and 10x downsampling, then fine-tune a ViT pretrained on ImageNet-1k with masked autoencoding. Using 5-fold cross-validation, they report a DICE score of 89.9%, precision 82.4%, recall 99.7%, and a Plexus Inclusion Rate (PIR) of 100%, and claim that this surpasses a CNN (89.2% DICE, 96.0% PIR) and k-means clustering (70.7% DICE, 77.4% PIR). The paper argues that perfect PIR is clinically important because excluded plexus regions would prevent downstream ganglion-cell detection.

Significance. If the reported comparison were conducted under strictly matched protocols, showing that a ViT reaches 100% PIR with only a modest DICE reduction would be a practically useful result for automating Hirschsprung's disease analysis, since PIR is the clinically relevant metric for downstream ganglion detection. The paper also proposes a clinically motivated evaluation metric and describes the preprocessing, training, and inference pipeline in reasonable detail. The ViT evaluation itself is an independent empirical measurement, not a circular fit. However, the strength of the central superiority claim is currently limited by the lack of matched-protocol baselines, the test-set-based threshold selection, and the absence of statistical uncertainty quantification.

major comments (4)

[Table 1 and Model Training and Testing] The CNN and k-means results in Table 1 are cited from references [5,6] rather than re-evaluated under the same protocol as the ViT. The manuscript states that the dataset is the same, but it does not demonstrate that the 5-fold split, tile size and stride, Macenko normalization, 10x downsampling, confidence threshold, and per-WSI aggregation are identical across methods. Since the ViT pipeline differs from the earlier CNN/k-means pipelines in at least the model and possibly the preprocessing details, the 0.7 percentage-point DICE gap and the 4 percentage-point PIR gap could be evaluation artifacts rather than evidence of model superiority. Please either re-run the baselines under the exact ViT evaluation protocol or clearly report the common protocol and justify that all comparison numbers are commensurate.
[Results and Figure 2] The operating point reported in the abstract and Table 1 (threshold 0.01 yielding DICE 89.9% and PIR 100%) was selected after inspecting the test-set threshold curve. This is a post-hoc choice of the confidence threshold on the same data used to evaluate performance, which makes the reported metrics optimistically biased and invalidates a clean comparison with the fixed-threshold CNN/k-means baselines. Please select the threshold on a validation set within each cross-validation fold (or use nested cross-validation), then report the test-set performance at that pre-specified threshold. If the threshold-sweep analysis is retained, it should be clearly labeled as an exploratory analysis and not the basis of the headline comparison.
[Results and Table 1] No measure of variability or statistical significance is provided. With 30 WSIs from 26 patients, the difference between 89.9% and 89.2% DICE, and between 100% and 96% PIR, may be within sampling noise. Please report per-WSI DICE and PIR distributions (e.g., mean ± standard deviation, median and interquartile range), confidence intervals via bootstrap, and paired significance tests such as the Wilcoxon signed-rank test for ViT versus CNN. Also report the total number of plexus regions per WSI to assess whether the 100% PIR is based on a very small number of regions and hence unstable.
[Abstract vs. Introduction and Table 1] The abstract cites the k-means DICE score as 80.7%, while both the Introduction and Table 1 report 70.7%. This is a direct numerical inconsistency that affects the paper's credibility. Please verify all reported numbers and ensure the abstract, Introduction, Results, and Table 1 are consistent. The same check should be applied to the PIR figure for k-means, which appears as 77.4% in all locations.

minor comments (6)

[Introduction] The phrase 'computation pathology' should be 'computational pathology'.
[Introduction] The phrase 'inter- and intra-rate variability' should be 'inter- and intra-rater variability'.
[References] Reference [7] attributes 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale' to Minderer et al.; this paper is by Dosovitskiy et al. (ICLR 2021). Please correct the citation, and also verify whether the intended reference is actually the ViT paper or a different Minderer et al. work.
[Figure 2] The figure caption does not identify which curve corresponds to DICE and which to PIR, and the axes are not labeled. Please add axis labels and a legend.
[Deep Learning Model] The ViT architecture is not fully specified. Please state the model variant (e.g., ViT-Base/16), number of transformer layers and heads, embedding dimension, and the number of parameters, since these details are necessary for reproducibility.
[Conclusion] The conclusion states that the ViT 'outperformed' the baselines without acknowledging the threshold-selection caveat and the lack of matched-protocol re-evaluation. Please temper the claim to match the evidence actually presented.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ViT result is an independent empirical evaluation, and reported baseline comparisons are external prior results rather than derivations from the paper's own fitted parameters.

full rationale

The paper's central contribution is an empirical segmentation result: a ViT is trained and tested with 5-fold cross-validation on 30 WSIs, and its DICE, precision, recall, and PIR are computed directly from predicted versus ground-truth masks using the stated equations (1) and (2). No predicted quantity is defined in terms of the fitted model outputs in a way that makes the evaluation tautological: the metrics are computed from the segmentation masks, and the threshold selection shown in Figure 2, while a test-set operating-point choice that could inflate reported performance, is not a case where the reported metric equals the fitted parameter by construction. The comparison to CNN and k-means baselines relies on numbers cited from the authors' prior work [5,6], but those are external published empirical results rather than assumptions built into the ViT derivation, and the ViT's own performance does not reduce to those citations. The inconsistencies in the reported k-means DICE (80.7% in the abstract versus 70.7% in Table 1 and the introduction) are correctness or reporting concerns, not circularity. Therefore, no load-bearing step in the derivation chain is equivalent to its own inputs, and the paper should not be penalized for circular reasoning.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper's claims rest heavily on the quality and comparability of manual annotations, the representativeness of the small dataset, the effectiveness of color normalization, and the comparability of prior baseline results. The only explicit fitted parameter is the confidence threshold selected post hoc from test data.

free parameters (1)

confidence_threshold = 0.01
Chosen post hoc from test results to achieve 100% PIR; it directly affects the reported DICE and precision.

assumptions (4)

domain assumption Ground truth manual segmentations by pathologists are accurate and consistent.
All metrics are computed against these annotations, and no inter-rater reliability is reported.
domain assumption The 30-WSI dataset is representative of the clinical population for Hirschsprung's disease.
The sample comes from a single institution and may not generalize to other scanners, stains, or patient populations.
domain assumption Macenko color normalization sufficiently removes staining variability without distorting tissue boundaries.
This preprocessing is used without validation of its effect on segmentation outcomes.
domain assumption Prior CNN and k-means results in [5,6] are directly comparable despite possible differences in evaluation protocol.
Table 1 mixes numbers from different studies, and no common evaluation protocol is described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Segmentation of Muscularis Propria in Colon Histopathology Images Using Vision Transformers for Hirschsprung's Disease." pith.science (2026). https://pith.science/paper/HWJAXWHE

@misc{pith2026241220571,
  author       = {Pith},
  title        = {Pith review of: Segmentation of Muscularis Propria in Colon Histopathology Images Using Vision Transformers for Hirschsprung's Disease},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HWJAXWHE}},
  note         = {Machine review of arXiv:2412.20571}
}

read the original abstract

Hirschsprung's disease (HD) is a congenital birth defect diagnosed by identifying the lack of ganglion cells within the colon's muscularis propria, specifically within the myenteric plexus regions. There may be advantages for quantitative assessments of histopathology images of the colon, such as counting the ganglion and assessing their spatial distribution; however, this would be time-intensive for pathologists, costly, and subject to inter- and intra-rater variability. Previous research has demonstrated the potential for deep learning approaches to automate histopathology image analysis, including segmentation of the muscularis propria using convolutional neural networks (CNNs). Recently, Vision Transformers (ViTs) have emerged as a powerful deep learning approach due to their self-attention. This study explores the application of ViTs for muscularis propria segmentation in calretinin-stained histopathology images and compares their performance to CNNs and shallow learning methods. The ViT model achieved a DICE score of 89.9% and Plexus Inclusion Rate (PIR) of 100%, surpassing the CNN (DICE score of 89.2%; PIR of 96.0%) and k-means clustering method (DICE score of 80.7%; PIR 77.4%). Results assert that ViTs are a promising tool for advancing HD-related image analysis.

Figures

Figures reproduced from arXiv: 2412.20571 by the authors.

**Figure 1.** Whole slide image of a cross-section of the colon. Zoomed-in portion shows a plexus region with ganglion cells indicated with red arrows [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Impact of confidence threshold on DICE score and plexus inclusion rate [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages

[6]

J. A. Kurian, Automated Identification of Myenteric Ganglia in Histopathology Images for the Study of Hirschsprung’s Disease, M.A.Sc. thesis, Carleton University, 2021

work page 2021
[1]

The placental distal villous hypoplasia pattern: interobserver agreement and automated fractal dimension as an objective metric,

A. Mukherjee et al., "The placental distal villous hypoplasia pattern: interobserver agreement and automated fractal dimension as an objective metric," Pediatric and Developmental Pathology, vol. 19, no. 1, pp. 31–36, 2016

work page 2016
[2]

Intraobserver and interobserver variation in the histopathological assessment of liver allograft rejection,

A. J. Demetris et al., "Intraobserver and interobserver variation in the histopathological assessment of liver allograft rejection," Hepatology, vol. 14, no. 5, pp. 751–755, 1991

work page 1991
[3]

Image Processing and Analysis of Histopathological Images Relating to Hirschsprung’s Disease,

J. Kurian et al., "Image Processing and Analysis of Histopathological Images Relating to Hirschsprung’s Disease," CMBES Proceedings , vol. 41, 2018

work page 2018
[4]

Color image processing in Hirschsprung's disease diagnosis,

M. T. K. Law, A. D. C. Chan, and D. El Demellawy, "Color image processing in Hirschsprung's disease diagnosis," in 2016 IEEE EMBS International Student Conference (ISC), 2016, pp. 1–4

work page 2016
[5]

Machine learning -based approach for fully automated segmentation of muscularis propria from histopathology images of intestinal specimens,

C. McKeen, F. Zabihollahy, J. Kurian, A. D. Chan, D. El Demellawy, and E. Ukwatta, "Machine learning -based approach for fully automated segmentation of muscularis propria from histopathology images of intestinal specimens," in Medical Imaging 2019: Digital Pathology, vol. 10956, pp. 146–151, Mar. 2019

work page 2019
[7]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” arXiv preprint arXiv:2010.11929, 2020. [Online]. Available: https://arxiv.org/abs/2010.11929

arXiv 2010
[8]

A method for normalizing histology slides for quantitative analysis,

M. Macenko et al. , "A method for normalizing histology slides for quantitative analysis," in 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2009, pp. 1107–1110

work page 2009

Show all 15 references

[9]

Applications of Transformers in Histopathological Image Analysis: A Comprehensive Survey,

A. Atabansi, et al., "Applications of Transformers in Histopathological Image Analysis: A Comprehensive Survey," Biomedical Engineering Online, vol. 23, no. 1, 2023. DOI: 10.1186/s12938-023-01157-0

2023 doi
[10]

CViTS -Net: A CNN - ViT Network with Skip Connections for Histopathology Image Classification,

A. Kanadath, J. A. A. Jothi, and S. Urolagin, "CViTS -Net: A CNN - ViT Network with Skip Connections for Histopathology Image Classification," IEEE Access, 2024

2024
[11]

CellViT: Vision Transformers for Precise Cell Segmentation and Classification,

L. Hörst, et al., "CellViT: Vision Transformers for Precise Cell Segmentation and Classification," arXiv preprint , 2023. [Online]. Available: https://arxiv.org/abs/2306.15350

2023 arXiv
[12]

Masked Autoencoders Are Scalable Vision Learners,

K. He et al., "Masked Autoencoders Are Scalable Vision Learners," Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2022, pp. 16000 –16009. doi: 10.1109/CVPR52688.2022.01554

2022
[13]

Deep learning in histopathology: the path to the clinic,

J. van der Laak, G. Litjens, and F. Ciompi, "Deep learning in histopathology: the path to the clinic," Nature Medicine, vol. 27, no. 5, pp. 775–784, 2021

2021
[14]

Hirschsprung disease: diagnosis and management,

J. Kessmann, "Hirschsprung disease: diagnosis and management," American Family Physician, vol. 74, no. 8, pp. 1319-1322, 2006

2006
[15]

Hirschsprung disease,

S. Lotfollahzadeh, M. Taherian, and S. Anand, "Hirschsprung disease," in StatPearls, StatPearls Publishing, 2023

2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [6]

J. A. Kurian, Automated Identification of Myenteric Ganglia in Histopathology Images for the Study of Hirschsprung’s Disease, M.A.Sc. thesis, Carleton University, 2021

work page 2021

[2] [1]

The placental distal villous hypoplasia pattern: interobserver agreement and automated fractal dimension as an objective metric,

A. Mukherjee et al., "The placental distal villous hypoplasia pattern: interobserver agreement and automated fractal dimension as an objective metric," Pediatric and Developmental Pathology, vol. 19, no. 1, pp. 31–36, 2016

work page 2016

[3] [2]

Intraobserver and interobserver variation in the histopathological assessment of liver allograft rejection,

A. J. Demetris et al., "Intraobserver and interobserver variation in the histopathological assessment of liver allograft rejection," Hepatology, vol. 14, no. 5, pp. 751–755, 1991

work page 1991

[4] [3]

Image Processing and Analysis of Histopathological Images Relating to Hirschsprung’s Disease,

J. Kurian et al., "Image Processing and Analysis of Histopathological Images Relating to Hirschsprung’s Disease," CMBES Proceedings , vol. 41, 2018

work page 2018

[5] [4]

Color image processing in Hirschsprung's disease diagnosis,

M. T. K. Law, A. D. C. Chan, and D. El Demellawy, "Color image processing in Hirschsprung's disease diagnosis," in 2016 IEEE EMBS International Student Conference (ISC), 2016, pp. 1–4

work page 2016

[6] [5]

Machine learning -based approach for fully automated segmentation of muscularis propria from histopathology images of intestinal specimens,

C. McKeen, F. Zabihollahy, J. Kurian, A. D. Chan, D. El Demellawy, and E. Ukwatta, "Machine learning -based approach for fully automated segmentation of muscularis propria from histopathology images of intestinal specimens," in Medical Imaging 2019: Digital Pathology, vol. 10956, pp. 146–151, Mar. 2019

work page 2019

[7] [7]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” arXiv preprint arXiv:2010.11929, 2020. [Online]. Available: https://arxiv.org/abs/2010.11929

arXiv 2010

[8] [8]

A method for normalizing histology slides for quantitative analysis,

M. Macenko et al. , "A method for normalizing histology slides for quantitative analysis," in 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2009, pp. 1107–1110

work page 2009