Pith. sign in

REVIEW 4 major objections 5 minor 18 references

PolypVision: A Three-Stage Hierarchical Deep Learning Framework for Classification and Segmentation of Colorectal Polyps

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Three-stage hierarchical deep learning can classify, segment, and subtype colorectal polyps in one pipeline, with about 0.99 AUC and 94.4% mAP@50 on Kvasir-SEG.

desk verdict A plausible three-stage polyp pipeline that currently only reports detection-level numbers; the histology AUC claim needs label provenance before it can be taken seriously. read the letter →

arxiv 2608.10649 v1 pith:F3MP547S submitted 2026-08-11 cs.CV

Hamidreza Bolhasani , Hamidreza Rastad , Amir Mohammad Akbari , Mohammad Tashakoripour , Parnian Asadollahi , Ata Khodami , Mojgan Forootan This is my paper · ORCID
classification cs.CV
keywords ColorectalpolypsDeeplearningEfficientNetV2-MUNet++PolypsegmentationclassificationGrad-CAMTransfer
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

PolypVision is a three-stage hierarchical deep learning framework intended to give a complete colorectal polyp assessment in one pipeline: first distinguishing adenomatous from hyperplastic polyps while also predicting Paris and JNet endoscopic patterns, then producing pixel-level segmentation along with a recommended resection method, and finally subtyping adenomas into tubular, tubulovillous, and villous. The authors report frame-classification AUC around 0.99 on Kvasir-SEG and CVC-ClinicDB, and a detection mAP@50 of 94.4% on Kvasir-SEG, matching or exceeding previous methods. The claim is that sharing one EfficientNetV2-M backbone across stages, with progressive transfer learning and task-specific losses, lets each task benefit from the others and yields clinically useful accuracy without hardware-specific adaptation. A reader should care because the pipeline follows the actual clinical sequence—detect, characterize, segment, resect, subtype—rather than solving each task in isolation, and the authors make it available as a freely usable web application.

What carries the argument

The load-bearing mechanism is the three-stage hierarchy itself, built on a shared EfficientNetV2-M backbone with 52.8 million parameters. Stage 1 fine-tunes this backbone with Focal Loss ($\gamma=2$, $\alpha=0.25$) for adenomatous/hyperplastic classification plus Paris and JNet heads; Stage 2 reuses the Stage 1 backbone as the UNet++ encoder and trains with Dice plus binary cross-entropy for segmentation and cross-entropy for resection recommendation; Stage 3 reinitializes the backbone from Stage 2 and uses Focal Loss with MixUp augmentation for three-way adenoma subtyping. Progressive transfer learning (Stage 1 to Stage 2 to Stage 3), layer-wise learning-rate decay, and early stopping on validation AUC or Dice carry the argument that limited labeled data can be leveraged across related clinical tasks.

What would settle it

Run the trained Stage 1 model on a dataset with explicit per-polyp histology labels—for example the ERCPMP pathology annotations the paper lists but never analyzes—and compute AUC for adenomatous versus hyperplastic on those ground-truth labels. If the AUC drops materially below ~0.99, or if no such histology labels were used in training, the classification claim is not supported by the reported experiments.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a single sequential pipeline—Stage 1 adenomatous-versus-hyperplastic classification with simultaneous Paris and JNet outputs, Stage 2 UNet++ segmentation with resection-method recommendation, and Stage 3 histologic subtyping—can reach state-of-the-art accuracy while reusing one shared backbone. The paper reports AUC of about 0.99 for frame classification on both Kvasir-SEG and CVC-ClinicDB and mAP@50 of 94.4% on Kvasir-SEG, and interprets Grad-CAM maps as evidence that the classifier attends to lesion surface texture, pit pattern, and margins. The intended implication is that hierarchical transfer learning across detection, segmentation, and classification is beneficial: each stage supplies the next with learned representations, and the combined framework covers the full clinical workflow in a device-independent way.

Load-bearing premise

The load-bearing premise is that the reported ~0.99 AUC actually measures adenoma-versus-hyperplastic classification, even though the two datasets used for that result contain pixel masks and no stated histology labels, and the paper never says where those labels come from.

Editorial extensions

If this is right

  • If the reported numbers hold, one installed pipeline can give an endoscopist the polyp type, Paris and JNet morphology, a boundary mask, a resection suggestion, and the adenoma subtype from a single image.
  • The mAP@50 of 94.4% on Kvasir-SEG places the detector between YOLO-LAN's 96.19% and YOLOv8-s's 91.16%, showing a segmentation-driven hierarchical approach can compete with dedicated detectors.
  • Because Stage 2's encoder is reused to initialize Stage 3, improvements in segmentation quality should flow directly into subtype classification, making progressive transfer a viable path for data-scarce medical imaging.
  • Device independence would mean colonoscopy units can adopt the framework without recalibrating to proprietary imaging modes or narrow-band systems.
  • Stage 1 errors propagate downstream by design: any adenoma misclassified as hyperplastic is never passed to the subtyping stage, so overall clinical accuracy is bounded by Stage 1 recall.

Reading between the lines

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

  • The headline AUC may be measuring something weaker than histological classification: Kvasir-SEG and CVC-ClinicDB provide only pixel masks and no stated histology labels, so unless the authors had separate pathology labels for those frames, the reported 'frame classification' may be polyp-versus-background detection rather than adenoma-versus-hyperplastic typing.
  • A direct test of the Stage 1 claim would be to evaluate the same trained model on the ERCPMP dataset's explicit pathology labels, which the paper lists in its datasets and data availability but never describes in the methods or results.
  • The Paris and JNet outputs and the resection-method recommendations are announced as part of the pipeline, yet no quantitative metrics are reported for them; their accuracy is an open empirical question.
  • A useful extension would be to report Stage 3 performance conditioned on Stage 1 being correct, isolating the hierarchical error propagation the authors acknowledge as a limitation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. PolypVision proposes a three-stage hierarchical deep learning framework for colorectal polyp analysis. Stage 1 performs binary classification of polyps as adenomatous or hyperplastic, together with Paris and JNet classification, using EfficientNetV2-M with Focal Loss. Stage 2 performs polyp segmentation with a UNet++ decoder, using the Stage 1 backbone as encoder, and also predicts a recommended resection method. Stage 3 classifies adenomas into tubular, tubulovillous, and villous subtypes using transfer learning from Stage 2. The paper reports an AUC of approximately 0.99 for frame classification and a detection mAP@50 of 94.4% on Kvasir-SEG, and includes a Grad-CAM interpretability analysis. The central technical idea of progressive transfer between classification, segmentation, and subtype classification is plausible, but the reported experiments do not substantiate most of the claimed outputs, and the main classification result appears to be measured on datasets that lack histology labels.

Significance. If the framework truly performed adenoma-versus-hyperplasia classification at AUC ~0.99 while simultaneously providing segmentation, Paris/JNet characterization, resection recommendations, and subtype classification, it would be a clinically valuable contribution. The use of public benchmarks, the shared-backbone transfer-learning design, and the publicly accessible web application are constructive elements. However, the manuscript as written does not provide evidence for the central classification claim, omits results for several advertised stages, and does not report the segmentation metrics promised in the evaluation protocol. The significance of the work therefore depends entirely on whether the authors can supply the missing dataset provenance and experimental results.

major comments (4)
  1. [§3, Table 1, §5.3] The Stage 1 adenoma/hyperplasia claim is unsupported by the reported experiments. The manuscript states in §3 that 'For Stage 1 (binary classification), images are labeled as adenomatous or hyperplastic,' but it never identifies which dataset provides these labels. Kvasir-SEG and CVC-ClinicDB are described as segmentation/detection datasets with pixel-wise and frame-level annotations, respectively, and contain no histopathology labels; PolypGen is likewise a detection/segmentation dataset. Table 1 lists ERCPMP as the only dataset with 'endoscopic and pathologic annotations,' yet no experiment in §5 evaluates ERCPMP. Consequently, the AUC ~0.99 for 'Frame Classification' in Table 3 cannot be interpreted as validation of the adenomatous-vs-hyperplastic classifier; it is at best a frame-level polyp-detection metric on datasets whose labels do not include polyp type. This is a load-bearing gap because Stage 2 and Stage 3 inherit the Stage 1 premise.
  2. [§5.1, §5.2, Table 2] The segmentation results that are central to Stage 2 are not reported. §5.1 states that Stage 2 is evaluated using Dice Score, IoU, and Precision/Recall, but the only quantitative result in §5.2 is mAP@50 for detection on Kvasir-SEG; no Dice or IoU values appear anywhere, and no CVC-ClinicDB result is given for PolypVision in Table 2. Moreover, the manuscript never defines how mAP@50 is computed for this model (e.g., whether bounding boxes are derived from predicted segmentation masks). Without segmentation metrics, the claimed UNet++ contribution and the subsequent transfer to Stage 3 cannot be assessed.
  3. [§4.3–§4.5, §5] Multiple components advertised in the abstract and method sections have no experimental evaluation. No results are reported for the Paris classification, JNet classification, recommended resection method (cold snare polypectomy, EMR, or ESD), or Stage 3 adenoma subtype classification (tubular, tubulovillous, villous). The only classification table (Table 3) reports generic 'Frame Classification' AUC on Kvasir-SEG and CVC-ClinicDB, datasets that contain none of these labels. These omissions are not cosmetic: the paper's title and abstract promise a complete clinical pipeline, yet the evidence covers only a subset of the claimed outputs.
  4. [§5.1, §5.3] Reported performance lacks statistical anchoring. The AUC values in Table 3 are given as '~0.99' without confidence intervals, error bars, or significance tests, and the number of test images per class is not reported. Given that Kvasir-SEG and CVC-ClinicDB are small datasets (1,000 and 612 images, respectively), the headline claim would be more credible with per-class counts and confidence intervals. Additionally, no confusion matrix or class-wise precision/recall is provided for the classification results.
minor comments (5)
  1. [Title page] The running header on the first page reads 'PolypVision: Ensemble Deep Learning for Colorectal Polyp Classification,' which does not match the main title describing a three-stage hierarchical framework; please harmonize the titles.
  2. [§3] The sentence 'Stratified train/validation/test splits are applied to preserve class balance across all partitions' is undefined for Stage 1 because the label source is not identified; once the label provenance is clarified, the split process should be described explicitly.
  3. [§5.4] The Grad-CAM analysis is only qualitative; a quantitative evaluation (e.g., overlap between attention heatmaps and polyp masks, or insertion/deletion metrics) would strengthen the interpretability claim.
  4. [§6] The device-independence claim is not supported by any acquisition-system or device metadata in the evaluation; multi-center data alone does not establish device independence.
  5. [Table 3] The citation for the ResNet-50/VGG16 AUC values is unclear; reference [12] is He et al. on ResNet, and the source of the 0.91–0.98 range should be cited precisely.

Circularity Check

2 steps flagged · score 6.0 of 10

Stage 1 histology AUC is either detection renamed or rests on an unevaluated self-authored dataset.

  1. self definitional [Section 3 (Datasets) and Section 5.3 (Table 3); also abstract]
    "For Stage 1 (binary classification), images are labeled as adenomatous or hyperplastic. ... Kvasir-SEG [16] provides 1,000 high-quality endoscopic polyp images with pixel-level segmentation masks ... CVC-ClinicDB [17] contains 612 frames extracted from 29 colonoscopy video sequences, providing frame-level annotations suited for segmentation and classification evaluation."

    Stage 1 is defined as a binary adenomatous-vs-hyperplastic classifier, but the reported evaluation is on Kvasir-SEG and CVC-ClinicDB, which Table 1 and Section 3 describe as providing only pixel-wise masks and frame-level annotations. No adenoma/hyperplasia labels are stated for either dataset, so the reported AUC ~0.99 cannot be a score for the Stage 1 histology classifier. Any AUC computed on these datasets must derive from the mask/frame labels available in the inputs, i.e., polyp-presence or segmentation labels; presenting it as 'frame classification' substitutes the input label modality for the claimed histological prediction. The headline classification result is therefore equivalent, by construction, to evaluating a detection-style label, not an independent histological classifier.

  2. self citation load bearing [Table 1 and Data Availability Statement; Ref [18]]
    "ERCPMP [18] Classification Image & Video 419 images, 37 videos Endoscopic and pathologic annotations Public 2024 ... All datasets used in this study are publicly available: ERCPMP [18] (https://doi.org/10.1186/s13104-024-07062-6)"

    ERCPMP is the only dataset in Table 1 carrying 'endoscopic and pathologic annotations,' and it is reference [18], whose authorship includes the present corresponding author, Mojgan Forootan. No experiment, table, or split in Section 5 uses ERCPMP, yet the Data Availability statement says it is 'used in this study.' If ERCPMP supplied the Stage 1 labels, the central histology-classification result depends entirely on a dataset created by the same authors and has no external validation; if it did not, there is no stated source for the adenoma/hyperplasia labels at all. Either branch makes the Stage 1 premise load-bearing on a self-citation rather than on independently verifiable data.

full rationale

The detection and segmentation results (mAP@50 = 94.4% on Kvasir-SEG, Dice/IoU on Kvasir-SEG and CVC-ClinicDB) are measured on public benchmarks and therefore are not circular. The circularity concern is confined to the Stage 1 classification claim. Table 3 presents 'frame classification AUC ~0.99' on Kvasir-SEG and CVC-ClinicDB as the Stage 1 result, but those datasets, as described in Table 1 and Section 3, supply only pixel masks and frame-level annotations; no adenomatous/hyperplastic labels are reported for them. Consequently, the reported AUC cannot be a score for the Stage 1 binary histology classifier defined in Section 4.3; if it was computed from mask/frame labels, it is a detection-style result renamed as classification, and if it was computed elsewhere, the label source is absent. The only dataset in the paper with pathologic annotations is ERCPMP, reference [18], whose authors overlap with the present corresponding author; it is listed as used but never evaluated in Section 5. That makes the classification premise depend, at best, on a self-authored dataset with no reported external validation. Because the Stage 1 claim is load-bearing for Stage 2 transfer and Stage 3 subtyping, the partial circularity warrants a score of 6; the independent detection and segmentation benchmarks on public data prevent a higher score.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The system's reported performance depends on several manual hyperparameters and on unverified assumptions about label provenance and metric comparability. The paper does not report ablations, sensitivity analyses, or error bars.

free parameters (5)
  • Focal Loss gamma = 2
    Set manually for class imbalance; no sensitivity analysis shown.
  • Focal Loss alpha = 0.25
    Set manually; no sensitivity analysis shown.
  • Weight decay (AdamW) = 1e-4
    Chosen manually; no ablation.
  • Learning rates = 1e-3 heads, 1e-4 to 5e-5 backbone
    Layer-wise decay set manually; no search reported.
  • Input resolution = 224x224 for classification
    Standard choice; no resolution study.
assumptions (3)
  • domain assumption Kvasir-SEG and CVC-ClinicDB contain labels for adenomatous vs hyperplastic polyp classification.
    These are segmentation datasets; the paper does not document any histology labeling for them.
  • domain assumption mAP@50 computed from segmentation masks is directly comparable to mAP@50 from object detectors.
    Table 2 compares against YOLO detectors, but no method is described for deriving detection boxes from the segmentation output.
  • domain assumption ERCPMP provides reliable ground truth labels for Stage 1 and Stage 3.
    ERCPMP is listed in the dataset table and references but never described in the methods; its creator is the corresponding author, creating an evaluation-on-own-data risk.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PolypVision: A Three-Stage Hierarchical Deep Learning Framework for Classification and Segmentation of Colorectal Polyps." pith.science (2026). https://pith.science/paper/F3MP547S

@misc{pith2026260810649,
  author       = {Pith},
  title        = {Pith review of: PolypVision: A Three-Stage Hierarchical Deep Learning Framework for Classification and Segmentation of Colorectal Polyps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F3MP547S}},
  note         = {Machine review of arXiv:2608.10649}
}
read the original abstract

Colorectal cancer (CRC) remains one of the leading causes of cancer-related mortality worldwide, predominantly arising from precancerous polyps. Accurate detection, segmentation, and endoscopic and histological classification of colorectal polyps are crucial for timely clinical intervention. In this study, we present PolypVision, a three-stage hierarchical deep learning framework that sequentially performs: (Stage 1) binary classification of polyps as adenomatous or hyperplastic, with simultaneous Paris and JNet classification, using EfficientNetV2-M with Focal Loss; (Stage 2) polyp segmentation with recommended resection method using a UNet++ decoder with the Stage 1 backbone as encoder, optimized with Dice and BCE losses; and (Stage 3) adenoma subtype classification (tubular, tubulovillous, villous) using EfficientNetV2-M with transfer learning from Stage 2. Evaluated on three public datasets -- PolypGen, Kvasir-SEG, and CVC-ClinicDB -- PolypVision achieves an AUC of approximately 0.99 for frame classification and a detection mAP@50 of 94.4% on Kvasir-SEG, outperforming or matching state-of-the-art methods. Gradient-weighted Class Activation Maps (Grad-CAM) confirm that the model attends to clinically relevant lesion features. The framework is device-independent, operating across diverse endoscopic imaging systems without hardware-specific adaptation. These results demonstrate that a hierarchical, transfer-learning-driven pipeline with task-specific loss functions offers a robust, device-independent, and clinically meaningful approach to automated colorectal polyp analysis. PolypVision is freely available as a web application at https://polypvision.com, a DataBioX initiative, with a free usage tier open to all users.

Figures

Figures reproduced from arXiv: 2608.10649 by the authors.

Figure 1
Figure 1. Representative image samples from the three datasets used in PolypVision: (A) PolypGen — multi-center endoscopic frames; (B) Kvasir-SEG — polyp images with pixel-wise segmentation masks; (C) CVC-ClinicDB — colonoscopy video frames [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. PolypVision three-stage hierarchical pipeline. Stage 1: binary classification (Adenomatous vs. Hyperplastic) using EfficientNetV2-M with Focal Loss. Stage 2: polyp segmentation using UNet++ with the Stage 1 backbone as encoder (Dice + BCE Loss). Stage 3: adenoma subtype classification (Tubular / Tubulovillous / Villous) using EfficientNetV2-M transferred from Stage 2, trained with Focal Loss and MixUp augmentation. … view at source ↗
Figure 3
Figure 3. Grad-CAM visualizations from the PolypVision Stage 1 classifier. Each panel pair shows the original endoscopic image (left) and the corresponding Grad-CAM heatmap (right). Warm colors (red/orange) indicate high model attention; cool colors (blue/green) indicate low activation. True Label 0 = hyperplastic; True Label 1 = adenomatous. The bottom-right panel illustrates a misclassified case (True Label 1, Predicted Cla… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages

  1. [1]

    Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries

    Bray F, Ferlay J, Soerjomataram I, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68(6):394 –424

    work page 2018

  2. [2]

    Cancer statistics, 2020

    Siegel RL, Miller KD, Jemal A. Cancer statistics, 2020. CA Cancer J Clin. 2020;70(1):7 –30

    work page 2020

  3. [3]

    Validation of a simple classification system for endoscopic diagnosis of small colorectal polyps using narrow-band imaging

    Iwatate M, Sano Y, Tanaka S, et al. Validation of a simple classification system for endoscopic diagnosis of small colorectal polyps using narrow-band imaging. Gastroenterology. 2013;143(3):612–621

    work page 2013

  4. [4]

    The American Society for Gastrointestinal Endoscopy PIVI on real -time endoscopic assessment of diminutive colorectal polyps

    Rex DK, Kahi CJ, O'Brien M, et al. The American Society for Gastrointestinal Endoscopy PIVI on real -time endoscopic assessment of diminutive colorectal polyps. Gastrointest Endosc. 2011;73(3):419 –422

    work page 2011

  5. [5]

    Guidelines for colonoscopy surveillance after screening and polypectomy

    Lieberman DA, Rex DK, Winawer SJ, et al. Guidelines for colonoscopy surveillance after screening and polypectomy. Gastroenterology. 2012;143(3):844–857

    work page 2012

  6. [6]

    WM-DOVA maps for accurate polyp highlighting in colonoscopy: Validation vs

    Bernal J, Sánchez FJ, Fernández-Esparrach G, et al. WM-DOVA maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput Med Imaging Graph. 2015;43:99 –111

    work page 2015

  7. [7]

    Deep learning localizes and identifies polyps in real time with 96% accuracy in screening colonoscopy

    Urban G, Tripathi P, Alkayali T, et al. Deep learning localizes and identifies polyps in real time with 96% accuracy in screening colonoscopy. Gastroenterology. 2018;155(4):1069–1078

    work page 2018

  8. [8]

    Automated classification of polyps using deep learning architectures and few-shot learning

    Tian P, He B, Qu Y, et al. Automated classification of polyps using deep learning architectures and few-shot learning. BMC Med Imaging. 2023;23(1):59

    work page 2023

Show all 18 references
  1. [9]

    EfficientNetV2: Smaller models and faster training

    Tan M, Le QV. EfficientNetV2: Smaller models and faster training. Proc ICML. 2021. arXiv:2104.00298

    2021 arXiv

  2. [10]

    UNet++: A nested U -net architecture for medical image segmentation

    Zhou Z, Rahman Siddiquee MM, Tajbakhsh N, Liang J. UNet++: A nested U -net architecture for medical image segmentation. Deep Learning in Medical Image Analysis. 2018:3–11

    2018

  3. [11]

    Ensemble learning: a survey

    Sagi O, Rokach L. Ensemble learning: a survey. WIREs Data Mining Knowl Discov. 2018;8(4):e1249

    2018

  4. [12]

    Deep residual learning for image recognition

    He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. Proc IEEE CVPR. 2016:770 –778

    2016

  5. [13]

    YOLOv8 for polyp detection: a comparative study

    Zhao Z, et al. YOLOv8 for polyp detection: a comparative study. arXiv. 2023

    2023

  6. [14]

    YOLO-LAN: A lightweight YOLO for polyp detection

    Wang C, et al. YOLO-LAN: A lightweight YOLO for polyp detection. arXiv. 2023

    2023

  7. [15]

    PolypGen: A multi-center polyp detection and segmentation dataset in colonoscopy

    Ali S, et al. PolypGen: A multi-center polyp detection and segmentation dataset in colonoscopy. arXiv:2106.04463. 2021

    2021 arXiv

  8. [16]

    Kvasir-SEG: A segmented polyp dataset

    Jha D, Smedsrud PH, Riegler MA, et al. Kvasir-SEG: A segmented polyp dataset. Proc MMM. 2020:451–462

    2020

  9. [17]

    CVC -ClinicDB: A clinical colonoscopy database for polyp detection

    Bernal J, Sánchez FJ, Fernández -Esparrach G, et al. CVC -ClinicDB: A clinical colonoscopy database for polyp detection. CVC Technical Report. 2015

    2015

  10. [18]

    ERCPMP: an endoscopic image and video dataset for colorectal polyps morphology and pathology

    Forootan, M., et al. ERCPMP: an endoscopic image and video dataset for colorectal polyps morphology and pathology. BMC Res Notes 17, 393 (2024). https://doi.org/10.1186/s13104 -024-07062-6 PolypVision: Ensemble Deep Learning for Colorectal Polyp Classification 10 Author Biogra...

    2024 doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.