pith. sign in

arxiv: 2605.15062 · v2 · pith:CW2OPNKGnew · submitted 2026-05-14 · 💻 cs.CV

Training-Time Optical Priors for Wireless Capsule Endoscopy Classification: Hemoglobin-Aware Input Fusion with Cross-Vendor Evaluation

Chengshuai Yang , Lei Xing , Keyaan Zawad Alam , Gregory Entin , Roopa Vemulapalli , Lisa Casey , Raiyan Tripti Zaman This is my paper

Pith reviewed 2026-06-30 21:14 UTC · model grok-4.3

classification 💻 cs.CV
keywords wireless capsule endoscopyhemoglobin prioroptical priorsinput fusionclassificationcross-vendor evaluationtraining-time priorsLymphangiectasia
0
0 comments X

The pith

A training-time hemoglobin optical prior fused with RGB inputs improves wireless capsule endoscopy classification, raising macro-AUC from 0.760 to 0.783.

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

The paper shows that standard RGB classifiers for wireless capsule endoscopy mix up hemoglobin signals with bile and lighting effects, hurting detection of vascular issues like Lymphangiectasia. It proposes injecting an analytic hemoglobin prior derived from light transport modeling only during training as an extra input channel. This fusion approach, tested on the Kvasir-Capsule dataset with patient-disjoint splits across multiple seeds and models, yields consistent gains in AUC, particularly for hard classes. A distillation version allows the benefit at inference without the extra channel. Cross-vendor tests on another dataset retain part of the gain.

Core claim

By feeding a Monte-Carlo-inspired hemoglobin prior P_blood alongside RGB channels into classifiers like EfficientNet-B0 during training, the method increases cross-seed macro-AUC from 0.760 to 0.783, with the three-stream model reaching 0.804; Lymphangiectasia AUC improves from 0.238 to 0.337 across all seeds, and gains hold in zero-shot transfer to the Galar cohort.

What carries the argument

The hemoglobin prior P_blood, an analytic approximation of light transport that isolates hemoglobin contrast, fused as an additional input channel at training time only.

If this is right

  • Input fusion with the prior improves macro-AUC and specific class performance like Lymphangiectasia.
  • Distillation allows RGB-only inference while retaining some gains.
  • Three-stream extensions combining spatial, temporal, and autoencoder streams further boost performance to 0.804 AUC.
  • Improvements are sign-consistent across seeds and replicate on ResNet-18 and ConvNeXt-Tiny.
  • Partial retention of gains in cross-vendor zero-shot transfer to Galar cohort.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach demonstrates that physics-based priors can be injected without increasing inference cost, potentially applicable to other medical imaging tasks where optical effects confound features.
  • Only the spatial-channel form of the prior provides benefit, suggesting the mechanism depends on explicit channel fusion rather than other parameterizations.
  • Future tests could check if similar priors for other tissue properties like bile would yield additive gains.
  • Patient-disjoint splits and multi-seed evaluation reduce the risk that gains are due to data leakage or random variation.

Load-bearing premise

The analytic Monte-Carlo-inspired hemoglobin prior accurately isolates hemoglobin contrast from bile staining and illumination falloff in the images and the classifier can use it effectively when added as a training-time input channel.

What would settle it

If applying the hemoglobin prior fusion on a new held-out WCE dataset shows no AUC improvement or if the prior fails to correlate with expert-labeled hemoglobin regions, the central claim would be falsified.

read the original abstract

Background. RGB-trained classifiers for wireless capsule endoscopy (WCE) conflate hemoglobin contrast with bile staining and illumination falloff, limiting sensitivity to small-vessel vascular findings such as Lymphangiectasia. We introduce a physics-informed framework that injects an analytic, Monte-Carlo-inspired hemoglobin prior into a standard classifier purely at training time -- to our knowledge the first use of an explicit optical light-transport prior in WCE classification. Methods. On Kvasir-Capsule (47,238 frames, 43 patients, 11 evaluable classes; patient-disjoint split) we test, across 6 seeds against an RGB-only EfficientNet-B0 baseline: (i) a 5-channel input-fusion variant feeding the prior P_blood alongside RGB; (ii) a distillation variant that runs on plain 3-channel RGB at inference; and (iii) a three-stream extension adding a temporal Transformer and an autoencoder-residual stream. We replicate across ResNet-18 and ConvNeXt-Tiny and report cross-vendor zero-shot transfer on the public Galar cohort. Results. Input fusion lifts cross-seed macro-AUC 0.760 -> 0.783 (5/6 seeds positive); distillation reaches 0.773; the three-stream model reaches 0.804 (+0.044 over baseline, paired DeLong p < 1e-4). Lymphangiectasia AUC rises 0.238 -> 0.337, sign-consistent across all 6 seeds. A four-variant ablation reveals a parameterization-mechanism boundary: only the spatial-channel form lifts. Cross-vendor zero-shot on Galar retains ~60% of the ConvNeXt-Tiny lift.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that an analytic Monte-Carlo-inspired hemoglobin prior P_blood, when fused as a fifth input channel at training time only, improves macro-AUC on the Kvasir-Capsule dataset (47k frames, patient-disjoint split) from 0.760 to 0.783 for EfficientNet-B0 (5/6 seeds), with larger gains for Lymphangiectasia (0.238 to 0.337); a three-stream extension reaches 0.804, distillation reaches 0.773, results replicate on ResNet-18/ConvNeXt-Tiny, and ~60% of the lift transfers zero-shot to the Galar cohort. A four-variant ablation indicates only the spatial-channel parameterization works.

Significance. If the prior is shown to isolate hemoglobin contrast, the training-time-only fusion approach would be a lightweight way to inject optical domain knowledge into WCE classifiers without changing inference, with potential value for vascular findings. The patient-disjoint splits, multi-seed reporting, DeLong tests, cross-architecture replication, and cross-vendor evaluation are strengths that support the empirical claims.

major comments (3)
  1. [Abstract/Methods] Abstract/Methods: The full derivation of the Monte-Carlo-inspired hemoglobin prior P_blood and the exact fusion equations are not supplied, preventing verification that the prior isolates hemoglobin contrast from bile staining and illumination falloff rather than acting as a generic extra channel.
  2. [Results] Results (four-variant ablation): The ablation demonstrates that only the spatial-channel form produces the reported AUC lift, but supplies no quantitative check (pixel-wise correlation with vascular masks, ROC against expert annotations, or controlled bile/illumination perturbation experiments) that P_blood performs the claimed optical separation on the 47k-frame Kvasir-Capsule images.
  3. [Results] Results (Lymphangiectasia and cross-seed): While the 0.238→0.337 AUC lift is sign-consistent across all 6 seeds, the absence of a mechanistic validation for P_blood means the improvement cannot yet be attributed to the optical prior rather than any informative fifth channel.
minor comments (2)
  1. [Abstract] The abstract states 'to our knowledge the first use' without a supporting literature comparison paragraph; a brief related-work sentence would clarify novelty.
  2. [Results] Cross-vendor Galar results are summarized as retaining '~60%' of the lift; reporting the exact retained delta and its statistical significance would strengthen the transfer claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback and accurate summary of our results. We agree that the derivation of P_blood requires explicit inclusion and will add it. For the mechanistic validation points, we provide context from the existing ablation and cross-vendor experiments while acknowledging the limits of the current study design.

read point-by-point responses

  1. Referee: [Abstract/Methods] Abstract/Methods: The full derivation of the Monte-Carlo-inspired hemoglobin prior P_blood and the exact fusion equations are not supplied, preventing verification that the prior isolates hemoglobin contrast from bile staining and illumination falloff rather than acting as a generic extra channel.

    Authors: We agree that the full derivation and fusion equations were omitted. In the revised manuscript we will add a dedicated Methods subsection presenting the complete Monte-Carlo light-transport derivation of P_blood (including hemoglobin absorption spectra, scattering parameters, and the closed-form approximation) together with the precise five-channel fusion equations used at training time. This will enable direct verification that the prior targets hemoglobin contrast. revision: yes

  2. Referee: [Results] Results (four-variant ablation): The ablation demonstrates that only the spatial-channel form produces the reported AUC lift, but supplies no quantitative check (pixel-wise correlation with vascular masks, ROC against expert annotations, or controlled bile/illumination perturbation experiments) that P_blood performs the claimed optical separation on the 47k-frame Kvasir-Capsule images.

    Authors: The referee correctly notes the absence of pixel-wise or perturbation-based optical validation. Kvasir-Capsule provides only classification labels and contains no vascular segmentation masks; controlled bile/illumination experiments would require a new acquisition protocol outside the scope of this work. We will add an explicit limitations paragraph discussing this gap. The four-variant ablation already shows that generic fifth-channel additions do not reproduce the gains, and the partial zero-shot retention on the independent Galar cohort supplies indirect support for optical specificity. revision: partial

  3. Referee: [Results] Results (Lymphangiectasia and cross-seed): While the 0.238→0.337 AUC lift is sign-consistent across all 6 seeds, the absence of a mechanistic validation for P_blood means the improvement cannot yet be attributed to the optical prior rather than any informative fifth channel.

    Authors: We acknowledge the attribution concern. The ablation was explicitly designed to test whether any fifth channel suffices; only the hemoglobin-inspired spatial-channel parameterization produced the reported lift. This boundary, together with consistent gains across three architectures and ~60% retention under cross-vendor shift, provides evidence against a purely generic-channel explanation. We will expand the discussion section to highlight these controls while noting that direct mechanistic imaging validation remains future work. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical gains measured on held-out splits

full rationale

The paper's central results are AUC lifts on patient-disjoint held-out Kvasir-Capsule splits and zero-shot Galar transfer. The hemoglobin prior P_blood is introduced as an analytic Monte-Carlo-inspired quantity injected at training time; no equation or ablation reduces the reported macro-AUC or per-class AUC values to quantities defined solely by fitted parameters or self-citations. Performance numbers remain independent of the prior's internal construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the validity of the Monte-Carlo-inspired hemoglobin prior and the assumption that its fusion at training time produces measurable gains on the stated dataset splits.

axioms (1)
  • domain assumption The Monte-Carlo simulation yields an analytic prior P_blood that isolates hemoglobin contrast from bile and illumination effects in WCE images.
    Invoked in the description of the physics-informed framework.
invented entities (1)
  • hemoglobin prior P_blood no independent evidence
    purpose: Provide optical hemoglobin information to the classifier during training only.
    Introduced as the core novel input; no independent falsifiable prediction outside the reported experiments is given in the abstract.

pith-pipeline@v0.9.1-grok · 5879 in / 1467 out tokens · 44677 ms · 2026-06-30T21:14:05.619246+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Application of Autofluorescence Endoscopy for Colorectal Cancer Screening: Rationale and an Update

    “Application of Autofluorescence Endoscopy for Colorectal Cancer Screening: Rationale and an Update. ” Gastroenterology Research and Practice 2012: 971383. https://doi.org/10.1155/2012/971383. DeLong, E. R., D. M. DeLong, and D. L. Clarke-Pearson

    work page doi:10.1155/2012/971383 2012

  2. [2]

    Comparing the Areas Under Two or More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach

    “Comparing the Areas Under Two or More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach. ” Biometrics 44: 837–45. https://doi.org/10.2307/2531595. Deng, J., W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei

    work page doi:10.2307/2531595

  3. [3]

    In2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255 (IEEE, 2009)

    “ImageNet: A Large- Scale Hierarchical Image Database. ” 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–55. https://doi.org/10.1109/CVPR.2009.5206848. Du, X., Y. Koronyo, N. Mirzaei, et al

    work page doi:10.1109/cvpr.2009.5206848 2009

  4. [4]

    Label-Free Hyperspectral Imaging and Deep-Learning 22 Prediction of Retinal Amyloid 𝛽-Protein and Phosphorylated Tau

    “Label-Free Hyperspectral Imaging and Deep-Learning 22 Prediction of Retinal Amyloid 𝛽-Protein and Phosphorylated Tau. ” PNAS Nexus 1 (4): pgac164. https://doi.org/10.1093/pnasnexus/pgac164. Habe, T. T., K. Haataja, and P. Toivanen

    work page doi:10.1093/pnasnexus/pgac164

  5. [5]

    Precision Enhancement in Wireless Cap- sule Endoscopy: A Novel Transformer-Based Approach for Real-Time Video Object Detection

    “Precision Enhancement in Wireless Cap- sule Endoscopy: A Novel Transformer-Based Approach for Real-Time Video Object Detection. ” Frontiers in Artificial Intelligence 8: 1529814. https://doi.org/10.3389/frai.2025.1529814. Houdeville, C., M. Souchaud, R. Leenhardt, et al

    work page doi:10.3389/frai.2025.1529814 2025

  6. [6]

    A Multisystem-Compatible Deep Learning-Based Algorithm for Detection and Characterization of Angiectasias in Small-Bowel Capsule Endoscopy: A Proof-of-Concept Study

    “A Multisystem-Compatible Deep Learning-Based Algorithm for Detection and Characterization of Angiectasias in Small-Bowel Capsule Endoscopy: A Proof-of-Concept Study. ” Digestive and Liver Disease 53 (12): 1627–31. https://doi.org/10.1016/j.dld.2021.08.026. Iddan, G., G. Meron, A. Glukhovsky, and P. Swain

    work page doi:10.1016/j.dld.2021.08.026 2021

  7. [7]

    Jacques, S

    https://doi.org/10.1038/35013140. Jacques, S. L

    work page doi:10.1038/35013140

  8. [8]

    Optical Properties of Biological Tissues: A Review

    “Optical Properties of Biological Tissues: A Review. ” Physics in Medicine and Biology 58: R37–61. https://doi.org/10.1088/0031-9155/58/11/R37. Kara, M. A., and J. J. Bergman

    work page doi:10.1088/0031-9155/58/11/r37

  9. [9]

    Autofluorescence Imaging and Narrow-Band Imaging for the Detection of Early Neoplasia in Patients with Barrett’s Esophagus

    “Autofluorescence Imaging and Narrow-Band Imaging for the Detection of Early Neoplasia in Patients with Barrett’s Esophagus. ” Endoscopy 38: 627–31. https://doi.org/10.1055/s-2006-925385. Liao, Z., R. Gao, C. Xu, and Z. S. Li

    work page doi:10.1055/s-2006-925385 2006

  10. [10]

    Indications and Detection, Completion, and Retention Rates of Small-Bowel Capsule Endoscopy: A Systematic Review

    “Indications and Detection, Completion, and Retention Rates of Small-Bowel Capsule Endoscopy: A Systematic Review. ” Gastrointestinal Endoscopy 71: 280–86. https://doi.org/10.1016/j.gie.2009.09.031. Loshchilov, I., and F. Hutter

    work page doi:10.1016/j.gie.2009.09.031 2009

  11. [11]

    EndoSLAM Dataset and an Unsupervised Monocular Visual Odometry and Depth Estimation Approach for Endoscopic Videos

    “EndoSLAM Dataset and an Unsupervised Monocular Visual Odometry and Depth Estimation Approach for Endoscopic Videos. ” Medical Image Analysis 71: 102058. https://doi.org/10.1016/j.media.2021.102058. Piccirelli, S., D. Salvi, C. L. Pugliano, et al

    work page doi:10.1016/j.media.2021.102058 2021

  12. [12]

    Pogorelov, K

    https://doi.org/10.3390/diagnostics150 91092. Pogorelov, K. et al

    work page doi:10.3390/diagnostics150

  13. [13]

    Bleeding Detection in Wireless Capsule Endoscopy Videos: Color Versus Texture Features

    “Bleeding Detection in Wireless Capsule Endoscopy Videos: Color Versus Texture Features. ” Journal of Applied Clinical Medical Physics 20: 141–54. https: //doi.org/10.1002/acm2.12662. Qi, D., S. Zhang, C. Yang, et al

    work page doi:10.1002/acm2.12662

  14. [14]

    Single-Shot Compressed Ultrafast Photography: A Review

    “Single-Shot Compressed Ultrafast Photography: A Review. ” Advanced Photonics 2 (1): 014003. https://doi.org/10.1117/1.AP.2.1.014003. Simula Research Laboratory

    work page doi:10.1117/1.ap.2.1.014003

  15. [15]

    Spada, C., S

    https://doi.org/10.1038/s41597-021-00920-z . Spada, C., S. Piccirelli, C. Hassan, et al

    work page doi:10.1038/s41597-021-00920-z

  16. [16]

    AI-Assisted Capsule Endoscopy Reading in Suspected Small Bowel Bleeding: A Multicentre Prospective Study

    “AI-Assisted Capsule Endoscopy Reading in Suspected Small Bowel Bleeding: A Multicentre Prospective Study. ” Lancet Digital Health 6 (5): e345–53. https://doi.org/10.1016/S2589-7500(24)00048-7. Sun, X., and W. Xu

    work page doi:10.1016/s2589-7500(24)00048-7

  17. [17]

    Fast Implementation of DeLong’s Algorithm for Comparing the Areas Under Correlated Receiver Operating Characteristic Curves

    “Fast Implementation of DeLong’s Algorithm for Comparing the Areas Under Correlated Receiver Operating Characteristic Curves. ” IEEE Signal Processing Letters 21: 1389–93. https://doi.org/10.1109/LSP.2014.2337313. Tan, M., and Q. V. Le

    work page doi:10.1109/lsp.2014.2337313 2014

  18. [18]

    Global, Regional, and National Burden of Early-Onset Gastric Cancer

    “Global, Regional, and National Burden of Early-Onset Gastric Cancer. ” Cancer Biology and Medicine 21: 667–78. https://doi.org/10.20892/j.issn.2095-3941.2024.0159. Wu, Z., C. Yang, X. Su, and X. Yuan

    work page doi:10.20892/j.issn.2095-3941.2024.0159 2095

  19. [19]

    Adaptive Deep PnP Algorithm for Video Snapshot Compressive Imaging

    “Adaptive Deep PnP Algorithm for Video Snapshot Compressive Imaging. ” International Journal of Computer Vision 131 (7): 1662–79. https: //doi.org/10.1007/s11263-023-01777-y . Yang, C., F. Cao, D. Qi, et al

    work page doi:10.1007/s11263-023-01777-y

  20. [20]

    Hyperspectrally Compressed Ultrafast Photography

    “Hyperspectrally Compressed Ultrafast Photography. ” Phys- ical Review Letters 124 (2): 023902. https://doi.org/10.1103/PhysRevLett.124.023902. Yang, C., S. Zhang, and X. Yuan

    work page doi:10.1103/physrevlett.124.023902

  21. [21]

    Ensemble Learning Priors Driven Deep Unfolding for Scalable Video Snapshot Compressive Imaging

    “Ensemble Learning Priors Driven Deep Unfolding for Scalable Video Snapshot Compressive Imaging. ” European Conference on Computer Vision (ECCV), 600–618. https://doi.org/10.1007/978-3-031-19790-1_36 . Zhao, R., C. Yang, R. T. Smith, and L. Gao

    work page doi:10.1007/978-3-031-19790-1_36

  22. [22]

    Coded Aperture Snapshot Spectral Imaging Fundus Camera

    “Coded Aperture Snapshot Spectral Imaging Fundus Camera. ” Scientific Reports 13 (1): 12007. https://doi.org/10.1038/s41598-023-39146- x. 24

    work page doi:10.1038/s41598-023-39146-