Generalizable Polyp Segmentation via Randomized Global Illumination Augmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate polyp segmentation in colonoscopy is crucial for colorectal cancer diagnosis.
• Deep learning models for polyp segmentation face domain shift challenges, reducing performance on new datasets.

Purpose of the Study:

• To develop an illumination enhancement-based domain generalization approach for polyp segmentation.
• To improve the generalization capability of deep learning models on unseen colonoscopy image datasets.

Main Methods:

• Proposed an image decomposition module (IDM) to separate images into reflectance and illumination components.
• Introduced an illumination transform module (ITM) for augmenting images with synthesized global illumination maps.
• Developed an illumination variance insensitiveness (IViSen) metric to assess model robustness.

Main Results:

• The proposed method demonstrated superior performance on unseen domains across four colonoscopy datasets (CVC-ClinicDB, CVC-ColonDB, ETIS-Larib, Kvasir-SEG).
• Achieved mean Dice of 60.82% and IoU of 53.19%, showing improvements of 2.06% and 2.31% respectively over competitive methods.
• The IViSen metric correlated well with model generalizability.

Conclusions:

• Illumination enhancement is an effective strategy to address domain shift in polyp segmentation.
• The proposed approach significantly improves the robustness and generalization of deep learning models for colonoscopy image analysis.
• This work offers a promising direction for enhancing automated polyp detection and colorectal cancer screening.