Computational Imaging Priors for Wireless Capsule Endoscopy: Monte Carlo-Guided Hemoglobin Mapping for Rare-Anomaly Detection

Technical Summary: Computational Imaging Priors for Wireless Capsule Endoscopy

Problem Statement
Current deep learning classifiers for Wireless Capsule Endoscopy (WCE), typically trained on standard RGB data, underperform on small-vessel vascular findings. These models often conflate hemoglobin contrast with confounding factors such as bile staining, intraluminal debris, and illumination falloff (specular highlights). While hardware solutions (e.g., spectral filters) exist, they are not universally deployed. The authors argue that existing AI treats capsule frames as generic photographs, ignoring the specific physics of the imaging system: the radially decaying fluence pattern of capsule LEDs and the distinct spectral absorption signature of hemoglobin.

Methodology
The study utilizes the Kvasir-Capsule dataset (47,238 frames, 11 evaluable classes) with a strict video-level 70/15/15 split to ensure generalization. The authors propose a Monte Carlo–inspired analytic prior to compute a hemoglobin probability map ($P_{blood}$) directly from standard RGB frames without additional hardware.

The core of the method involves three analytic steps:

1. Hemoglobin Index ($H$): A per-pixel ratio $R/(G+B)$ biased toward regions where hemoglobin absorbs green/blue light.
2. Robust Normalization: To prevent vignette dark pixels and saturated specular highlights from collapsing the dynamic range, the index is clipped to the 1st–99th percentile per frame before min-max normalization.
3. Radial Fluence ($\Phi$): A first-order approximation of the LED light transport, modeled as an exponential decay from the image center.
4. Prior Synthesis: The final prior is $P_{blood} = \sigma(\alpha(H_{norm} - 0.5)) \cdot \Phi(r)$.

Two software configurations are evaluated against an RGB-only EfficientNet-B0 baseline across 6 random seeds:

• Input-Fusion (5-Channel): The 3-channel RGB is concatenated with two physics channels ($P_{blood}$ and an autofluorescence surrogate $H_{AFI}$). The new channels are zero-initialized, ensuring the model behaves identically to the RGB baseline at initialization.
• Distillation (3-Channel + Auxiliary Head): A 3-channel RGB backbone is trained with an auxiliary decoder to predict the analytic $P_{blood}$ map. At inference, the decoder is discarded, leaving a standard 3-channel model that has internalized hemoglobin-aware representations.

Key Contributions

1. Analytic Prior Integration: The paper demonstrates that a closed-form, training-free physics prior can be fused into deep learning pipelines via zero-initialized auxiliary channels, improving sensitivity to subtle pathology without hardware changes.
2. Deployment-Friendly Distillation: A pathway is established where the physics prior acts as a teacher signal during training but is removed at inference. This allows the deployed model to run on standard 3-channel RGB inputs while retaining the benefits of the physics-informed feature space.
3. Methodological Insights:
   • Percentile Clipping: The authors identify that per-frame percentile clipping is essential; without it, outliers cause the prior to collapse, and the model learns to ignore the auxiliary channels.
   • Seed Variance: The study highlights that rare focal-vascular classes (e.g., Angiectasia) exhibit high variance across seeds, cautioning against over-interpreting single-seed results.

Results
Evaluated on the Kvasir-Capsule test split ($n=6,423$ frames):

• Macro-AUC: The Input-Fusion variant achieved a direction-consistent improvement over the RGB-only baseline, increasing macro-AUC from 0.760 ± 0.027 to 0.783 ± 0.024 (paired $\Delta = +0.023$, significant on 5/6 seeds). The Distillation variant achieved a similar mean (0.773 ± 0.028) but showed mixed directionality across seeds (3/6 positive).
• Per-Class Performance: The most robust improvement was observed on Lymphangiectasia, where the Input-Fusion variant improved AUC from 0.238 to 0.337 (sign-consistent across all 6 seeds).
• High-Variance Cases: A single seed (seed=42) showed a dramatic jump in Angiectasia AUC (0.528 to 0.916), but the cross-seed mean showed no consistent gain (0.646 to 0.608), illustrating the dataset's sensitivity to initialization for rare classes.
• Bleeding Detection: The "Blood - fresh" class showed no consistent gain, which the authors attribute to data scarcity (only 22 training frames from a single video) rather than methodological failure.

Significance and Claims
The paper claims that a Monte Carlo–inspired analytic prior provides a small but direction-consistent macro-AUC improvement on the Kvasir-Capsule benchmark. The authors emphasize that the distillation variant is the preferred deployment strategy: it matches the performance of the input-fusion variant while running on standard 3-channel RGB inputs, avoiding integration costs in commercial pipelines.

The work is modest in its claims, explicitly noting that:

• The gains are small in absolute terms but reproducible in direction.
• The method does not solve the "Blood - fresh" detection problem due to data limitations.
• Single-seed results on rare classes (like Angiectasia) should not be cited as typical performance due to high variance.
• The findings are currently limited to the Kvasir-Capsule dataset and require prospective clinical validation.

The authors conclude that integrating analytic optical priors with deep learning is a viable strategy for improving WCE classification, particularly for rare vascular anomalies, without requiring new hardware. All code and weights are released to ensure reproducibility.