PatchCL-AE: Anomaly detection for medical images using patch-wise contrastive learning-based auto-encoder.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Anomaly detection in medical imaging is crucial but challenging.
• Reconstruction-based methods are limited by pixel-level loss reliance.
• Existing methods struggle with accurate localization of anomalies.

Purpose of the Study:

• To develop an advanced medical anomaly detection method.
• To overcome limitations of reconstruction-based approaches.
• To improve the accuracy and localization of detected anomalies.

Main Methods:

• Proposed a patch-wise contrastive learning-based auto-encoder.
• Introduced a patch-wise contrastive learning loss for local semantic supervision.
• Designed an anomaly score based on local semantic discrepancies.

Main Results:

• Achieved state-of-the-art performance on three public datasets (brain MRI, retinal OCT, chest X-ray).
• Demonstrated over 99% Area Under the Curve (AUC) on retinal and brain images.
• The method effectively learns local normal features and improves discriminability of anomalous regions.

Conclusions:

• Patch-wise contrastive supervision enhances learning of local semantics.
• The patch-discrepancy score accurately pinpoints abnormalities.
• The proposed method offers targeted advancements for medical anomaly detection.