Self-supervised Anomaly Detection with Random-shape Pseudo-outliers.

Area of Science:

• Medical Imaging Analysis
• Machine Learning
• Computer Vision

Background:

• Anomaly detection in medical images is crucial but challenging due to the lack of abnormal samples during training.
• Existing methods often struggle with realistic anomaly representation, relying on simplified geometric shapes.

Purpose of the Study:

• To propose a novel self-supervised learning method for precise anomaly detection and localization in MRI.
• To develop an outlier synthesis strategy capable of generating random-shape anomalies for robust training.

Main Methods:

• A self-supervised learning framework utilizing a discriminative model for anomaly detection.
• A novel outlier synthesis strategy to generate pseudo-abnormalities with random shapes, mimicking natural anomalies.
• Learning the disentanglement of normal and pseudo-outlier regions within synthesized MRI images.

Main Results:

• The proposed method effectively detects and localizes anomalies at both pixel and sample levels.
• Empirical evaluation on two public MRI datasets demonstrates superior performance compared to state-of-the-art (SOTA) solutions.
• The random-shape anomaly synthesis proves more effective than pre-determined geometric shapes.

Conclusions:

• The novel self-supervised approach with random-shape anomaly synthesis significantly advances MRI anomaly detection.
• This method offers a robust solution for scenarios with limited or no abnormal training data.
• The technique shows promise for improving diagnostic accuracy in medical imaging.