Computational Imaging Priors for Wireless Capsule Endoscopy: Monte Carlo–Guided Hemoglobin Mapping for Rare-Anomaly Detection
2026
Read Computational Imaging Priors for Wireless Capsule Endoscopy: Monte Carlo–Guided Hemoglobin Mapping for Rare-Anomaly Detection (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-15062.txt.
Introduction
Chengshuai Yang and six colleagues inject a hemoglobin-aware optical prior into wireless capsule endoscopy classification at training time only. Standard RGB classifiers confuse hemoglobin contrast with bile staining and lighting falloff, missing small-vessel lesions. The prior, inspired by Monte Carlo light transport, lifts cross-seed macro-AUC from 0.760 to 0.783 under input fusion and to 0.773 under distillation, with a three-stream model reaching 0.804. Cross-vendor transfer on the Galar cohort is also assessed.
Catalog record
| YEAR | 2026 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63084 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.15062
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Computational Imaging Priors for Wireless Capsule Endoscopy: Monte Carlo–Guided Hemoglobin Mapping for Rare-Anomaly Detection Abstract Chengshuai Yang 1 , Lei Xing 3 , Gregory Entin 4 , Roopa Vemulapalli 5 , Lisa Casey 6 , Raiyan Tripti Zaman 1,2,∗ 1 Department of Biomedical Engineering, University of Texas Southwestern Medical Center, Dallas, TX 75390, USA 2 Department of Internal Medicine, University of Texas Southwestern Medical Center, Dallas, TX 75390, USA 3 Department of Radiation Oncology, Stanford University, Stanford, CA 94305, USA 4 VELVETECH, LLC, www.velvetech.com , USA — Chief Technology Officer (G.E.) 5 Division of Digestive and Liver Diseases, Clements University Hospital, Dallas, TX, USA — Clinical Chief and Director of Endoscopy Operations (R.V.) 6 Internal Medicine, Division of Digestive and Liver Diseases, Parkland Hospital, Dallas, TX, USA (L.C.) ∗ Correspondence: [email protected] (R.T.Z.) Article type: Article. Preprint of a manuscript in preparation for journal submission. Abstract Background. RGB-trained capsule-endoscopy classifiers underperform on small-vessel vascular findings by conflating hemoglobin contrast with bile and illumination falloff. Thus, here we test whether a Monte Carlo–inspired analytic model can compute hemoglobin from RGB signal built upon extracted classifier. Methods. On Kvasir-Capsule (47,238 frames, video-level 70/15/15 split, 11 evaluable classes) we evaluate two software-only configurations against RGB-only EfficientNet-B0 across 6 seeds: (i) a prior P_{\mathrm{blood}}=\sigma(\alpha(H_{\mathrm{norm}}-0.5))\cdot\Phi(r) fused as 2 zero-init auxiliary channels; (ii) a distillation head training a 3-channel RGB backbone to predict P_{\mathrm{blood}} . Significance: paired DeLong, McNemar, bootstrap CIs with Bonferroni correction. Results. Across 6 seeds ( n=6{,}423 ), macro-AUC is RGB 0.760\pm 0.027 , input-fusion 0.783\pm 0.024 (paired \Delta=+0.023 , sign-positive on 5/6 seeds), distillation 0.773\pm 0.028 (paired \Delta=+0.013 , sign-mixed at 3/6 seeds positive). The largest robust per-class lift is on Lymphangiectasia: RGB 0.238\pm 0.057 \to input-fusion 0.337\pm 0.019 \to distillation 0.334\pm 0.083 , sign-consistent across all 6 seeds. Per-class AUCs on rare focal-vascular classes (Angiectasia, Blood - fresh) show large cross-seed variance ( \sigma_{\mathrm{PI}}=0.23 on Angiectasia); seed=42 alone reaches Angiectasia AUC 0.528\to 0.916 , but the cross-seed mean is 0.646\to 0.608 , so we report this as a high-variance per-seed exemplar. Conclusion. A Monte Carlo–inspired analytic prior provides a small, direction-consistent macro-AUC improvement on Kvasir-Capsule across 6 seeds with the largest robust per-class lift on Lymphangiectasia; the distillation variant’s macro-AUC matches input-fusion’s at modestly higher cross-seed variance and runs on plain 3-channel RGB at inference, which is its deployment advantage. Keywords: capsule endoscopy; deep learning; Monte Carlo light transport; physics-informed neural networks; hemoglobin imaging; gastrointestinal cancer; small-bowel bleeding; Kvasir-Capsule 1. Introduction Gastrointestinal (GI) cancers — colorectal, esophageal, gastric, and small-bowel — collectively account for approximately 3.4 million deaths annually worldwide (Tan et al. 2024). Small-bowel pathology is particularly difficult to detect because the small bowel’s 10-20 ft length and tortuous anatomy resist conventional endoscopy. Wireless capsule endoscopy (WCE), introduced clinically in 2000 (Iddan et al. 2000), enables passive imaging of the small intestine at acquisition rates of 2-6 frames per second over 8-12 hours. The clinical workflow remains, however, fundamentally bottlenecked by manual review: a single 8-hour study yields ~50,000 frames, of which only a small minority contain actionable findings. Inter-reader sensitivity for small-bowel bleeding ranges from 56% to 89% depending on reader expertise, and 6-7 hours of expert review per study is typical (Liao et al. 2010). Three data-side limitations of current commercial PillCam® systems compound the burden: passive broadband white-light imaging that lacks biologically specific contrast for early dysplasia (Kara and Bergman 2006); coarse spatial localization that relies on transit-time heuristics; and a high false-positive rate when standard RGB classifiers are confounded by bile staining, intraluminal debris, or specular highlights. The first two of these limitations are best addressed by hardware (autofluorescence excitation, spectral filters, ultra-wideband ranging) (Aihara et al. 2012); the third can be substantially mitigated in software . Existing AI for capsule endoscopy is dominated by single-task RGB classifiers — polyp detection (Pogorelov et al. 2019), bleeding detection from color and texture features (Pogorelov et al. 2019), or video-level anomaly screening — that treat the captured frame as a generic photograph. Recent prospective clinical trials show that AI-assisted reading achieves non-inferior diagnostic yield for small-bowel bleeding while reducing reading time from 33.7 to 3.8 minutes per study (Spada et al. 2024), and recent reviews (Piccirelli et al. 2025) highlight remaining unmet needs in calibrated bleeding/vascular detection. Transformer-based RT-DETR variants now reach near-perfect F1 score — the harmonic mean of precision and recall, \text{F1}=2\cdot\frac{\text{precision}\cdot\text{recall}}{\text{precision}+\text{recall}} — on bleeding and angiectasia classes on Kvasir-Capsule (Habe et al. 2025), and dedicated vascular-lesion CNNs have been validated multi-system across PillCam and MiroCam capsules (Houdeville et al. 2021). All of this work treats the captured frame as a generic photograph. We argue that a physics-aware representation, even a simple one derived analytically from the same RGB data, can both improve sensitivity to subtle pathology and reduce false positives caused by illumination geometry. Specifically, capsule LEDs produce a centrally peaked, radially decaying fluence pattern that is well approximated by a first-order Monte Carlo light-transport model; and hemoglobin’s preferential absorption of green and blue light gives blood-rich regions a quantitatively distinct R/(G+B) signature that the network can exploit if it is given as an explicit input channel. The pairing of analytic optical priors with deep learning has precedent in adjacent medical-imaging domains — for instance, label-free hyperspectral imaging combined with deep learning for retinal amyloid \beta and tau prediction (Du et al. 2022) and snapshot spectral fundus imaging (Zhao et al. 2023) — and we extend that line of reasoning to single-channel-RGB capsule endoscopy. This paper makes three contributions: 1. A direction-consistent macro-AUC gain from an analytic hemoglobin prior, robust on the input-fusion variant. Across 6 independent seeds on Kvasir-Capsule’s video-level test split, the input-fusion variant improves macro-AUC — the unweighted mean of one-vs-rest area under the receiver operating characteristic curve (AUC) across the 11 evaluable classes — by paired \Delta=+0.023 (sign-positive on 5 of 6 seeds), from RGB-only at 0.760\pm 0.027 to 0.783\pm 0.024 . The distillation variant matches input-fusion’s expected macro-AUC ( 0.773\pm 0.028 , paired \Delta=+0.013 ) but is direction-mixed across seeds (3/6 sign-positive). The prior is computable at zero training cost from any standard RGB capsule frame, requires no additional capsule hardware, and slots into any CNN or transformer backbone via a 5-channel input expansion with zero-initialized weights for the physics channels. 2. A deployment-friendly distillation pathway. Distillation trains a 3-channel-RGB backbone to predict the analytic P_{\mathrm{blood}} map from spatial features (CE + \lambda \cdot BCE), discardable at deployment via model.deploy_mode = True . The deployed model is a plain RGB classifier with hemoglobin-aware internalized representations — addressing the integration cost of input-channel modifications in commercial PillCam® pipelines and producing a free per-frame interpretability heatmap. The robust per-class lift across all 6 seeds is on Lymphangiectasia: AUC improves 0.238\pm 0.057\to 0.337\pm 0.019\to 0.334\pm 0.083 across the three arms (sign-consistent on all 6 seeds for the input-fusion variant, with the lowest cross-seed \sigma in the table — 0.019 ). The single-seed=42 result of Angiectasia AUC 0.528\to 0.916 is preserved as a high-variance per-seed exemplar (cross-seed mean 0.646\to 0.608 , \sigma_{\mathrm{PI}}=0.23 ); we use it to illustrate the prior’s potential on hemoglobin-dominant lesions but do not present it as a typical performance. 3. Two diagnostic findings about analytic-prior design. (i) The prior’s per-frame percentile