Abstract
Domain Generalization (DG) is a critical challenge in Gastrointestinal (GI) endoscopy image analysis. Current open-source datasets are often limited by hardware-specific biases, as they are typically collected from a small number of patients using a single brand of sensor. These variations in device hardware and individual patient physiology significantly hinder the generalizability of AI models. To address this, we propose a multi-domain GI dataset and a lightweight, plug-in module called InvNorm (Invertible Normalization). Unlike previous DG methods that rely on non-invertible "black box" transformations, InvNorm utilizes normalizing flows to achieve invertible and explainable style normalization. Our method demonstrates superior performance across GI recognition, object detection, and standard natural image benchmarks.
The Challenge of Domain Gaps in Endoscopy
In clinical GI endoscopy, AI models frequently encounter a "domain gap." A model trained on images from one manufacturer's capsule camera may fail when deployed on another brand due to differences in lighting, resolution, and color rendering.
Existing DG techniques attempt to bridge this gap by forcing models to ignore "style" features and focus on "content." However, many of these methods are non-invertible, meaning they can introduce misleading or "hallucinated" features during data augmentation. In medical settings, this lack of transparency raises significant ethical concerns regarding the reliability of AI-driven diagnoses.
The InvNorm Architecture
The InvNorm module is designed as a lightweight "plug-in" that can be integrated at the beginning of any standard neural network backbone, such as ResNet.
Key Components:
- Invertible Encoding: Using a Normalizing Flow architecture (based on the GLOW framework), the module maps input images into a latent feature space. This bijective mapping ensures the process is mathematically reversible.
- Instance Normalization (IN): Within the latent space, Instance Normalization shifts features from disparate domains into a shared, stable style domain.
- Explainability through Invertibility: By ensuring the transformation is invertible, InvNorm guarantees that normalized features still correspond to a valid image space. This allows for the theoretical mapping of features back into viewable images, providing a layer of clinical explainability.
- Efficiency: The module utilizes "squeeze" operations for multi-scale processing and flow layers (including Actnorm and coupling layers) to transform data without information loss. Adding InvNorm to a ResNet-18 backbone increases the parameter count by only ~4M (from 11.7M to 15.7M).
Experimental Evaluation
The effectiveness of InvNorm was validated using two primary datasets:
- Multi-domain GI Dataset: A newly introduced dataset featuring four lesion categories—Vascular, Ulceration, Erosion, and Bleeding—collected from ten different sensors to test cross-hardware robustness.
- PACS Benchmark: A standard DG dataset (Art Painting, Cartoon, Photo, Sketch) used to evaluate performance on natural images.
Key Results
InvNorm significantly outperformed state-of-the-art (SOTA) methods like MixStyle:
- GI Object Detection: Achieved an AP50 of 0.217, compared to 0.171 for MixStyle.
- GI Image Recognition: Reached 48.29% accuracy, a substantial improvement over MixStyle’s 35.83%.
- PACS Benchmark: Set a new SOTA with an average accuracy of 84.2%.
Conclusion and Future Directions
InvNorm provides a robust framework for domain generalization in medical imaging by prioritizing mathematical invertibility. This approach not only improves performance across diverse hardware but also addresses the ethical necessity for explainable AI in healthcare. Future work will focus on optimizing the module for real-time inference on low-power capsule endoscopy receivers and conducting live clinical validations to assess performance in real-world "out-of-distribution" scenarios.
Summary updated