PolypMixNet: Enhancing semi-supervised polyp segmentation with polyp-aware augmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• AI-assisted polyp segmentation is vital for early colorectal cancer (CRC) detection.
• Supervised learning for polyp segmentation is hindered by insufficient annotated data.
• Existing semi-supervised methods struggle with class imbalance and task-specific challenges.

Purpose of the Study:

• To develop an effective semi-supervised learning framework for accurate colonoscopic polyp segmentation.
• To overcome limitations of scarce annotated data and class imbalance in polyp detection.
• To enhance diagnostic capabilities for colorectal cancer through improved segmentation.

Main Methods:

• Proposed PolypMixNet, a semi-supervised framework using a Mean Teacher architecture.
• Introduced polyp-aware mixup (PolypMix) for data augmentation and class balance.
• Implemented polyp-directed soft pseudo-labeling (PDSPL) and dual-level consistency regularization (PMPC, PMAC).

Main Results:

• Achieved 88.97% Dice and 88.85% mIoU on Kvasir-SEG dataset.
• Outperformed state-of-the-art semi-supervised methods by 2.88% Dice with only 15% labeled data.
• Demonstrated performance comparable to fully supervised methods.

Conclusions:

• PolypMixNet successfully tackles limited data and class imbalance in polyp segmentation.
• The framework leverages unlabeled data with novel augmentation and consistency techniques.
• This work advances semi-supervised learning in medical imaging for improved CRC diagnosis.