SSL-CPCD: Self-Supervised Learning With Composite Pretext-Class Discrimination for Improved Generalisability in...

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Supervised deep learning methods in medical imaging require extensive data and struggle with generalizability due to high variability in endoscopic data.
• Existing self-supervised learning (SSL) methods show promise but have a performance gap in the medical domain.

Purpose of the Study:

• To develop a novel SSL approach for medical image analysis that enhances feature representation and improves generalizability.
• To address the challenges of inter- and intra-patient variability in endoscopic imaging.

Main Methods:

• Proposed a patch-level instance-group discrimination method with additive angular margin penalization within cosine similarity metrics.
• This approach encourages learning clustered representations and better class separation.

Main Results:

• Achieved significant improvements over state-of-the-art (SOTA) methods on classification, detection, and segmentation tasks.
• Demonstrated notable performance: 79.77% accuracy for ulcerative colitis classification, 88.62% mAP for detection, and 82.32% Dice score for polyp segmentation.
• Showcased superior generalizability to unseen datasets with over 7% improvement compared to SOTA.

Conclusions:

• The proposed SSL method (SSL-CPCD) effectively learns robust feature representations from medical images, overcoming limitations of traditional supervised methods.
• The approach significantly enhances model performance and generalizability across various medical imaging tasks, facilitating clinical translation.