Self-supervised pseudo multi-class pre-training for unsupervised anomaly detection and segmentation in medical images

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Unsupervised anomaly detection (UAD) is crucial for medical image analysis (MIA) when training data predominantly contains normal images.
• Current UAD methods struggle with learning effective representations, leading to poor detection of diverse abnormalities.
• Existing self-supervised learning pre-training methods for UAD are suboptimal, lacking domain knowledge integration and effective normal data clustering.

Purpose of the Study:

• To propose a novel self-supervised pre-training method, PMSACL, to enhance UAD in MIA.
• To improve the sensitivity of UAD methods to detect and segment unseen abnormal lesions.
• To address the limitations of existing pre-training strategies by incorporating domain knowledge and promoting dense feature clusters.

Main Methods:

• Developed Pseudo Multi-class Strong Augmentation via Contrastive Learning (PMSACL), a new pre-training approach.
• PMSACL employs an optimization strategy that contrasts normal images against synthesized abnormal pseudo-classes.
• Each class (normal and pseudo abnormal) is trained to form a dense cluster in the feature space.

Main Results:

• PMSACL pre-training significantly improves the accuracy of state-of-the-art UAD methods.
• Demonstrated effectiveness across multiple MIA benchmarks, including colonoscopy, fundus screening, and COVID-19 Chest X-ray datasets.
• The proposed method enhances the ability of UAD to detect and segment various types of abnormalities.

Conclusions:

• PMSACL offers a superior self-supervised pre-training strategy for UAD in medical imaging.
• The method effectively addresses the challenge of learning robust image representations from normal-only training data.
• PMSACL shows broad applicability and improved performance on diverse medical imaging datasets.