Non-equivalent images and pixels: Confidence-aware resampling with meta-learning mixup for polyp segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels at polyp segmentation for colorectal cancer diagnosis but requires extensive pixel-level annotations.
• Existing models struggle with polyp variability, leading to unstable training.
• Limited annotated data hinders the development of robust polyp segmentation models.

Purpose of the Study:

• To develop advanced data augmentation and training strategies for more accurate and stable polyp segmentation.
• To address the challenges of limited annotations and inherent variability in polyp images.
• To improve the performance of deep learning models in segmenting polyps from endoscopic images.

Main Methods:

• Proposed Meta-Learning Mixup (MLMix) for adaptive data augmentation, converting soft labels to hard labels and expanding training datasets.
• Introduced Confidence-Aware Resampling (CAR) to progressively select confident samples and pixels, enhancing model representation and training stability.
• Implemented CAR with class distribution prior knowledge to rebalance polyp and normal class data.

Main Results:

• The proposed MLMix and CAR strategies significantly improved polyp segmentation performance.
• Achieved state-of-the-art results with 87.450% dice on the EndoScene test set.
• Demonstrated high performance with 86.453% dice on the wireless capsule endoscopy (WCE) polyp dataset.

Conclusions:

• MLMix and CAR are effective methods for enhancing polyp segmentation in endoscopic imaging.
• The developed approach overcomes limitations of laborious annotations and training instability.
• This work contributes to more reliable automated polyp detection for colorectal cancer management.