Dual-distribution discrepancy with self-supervised refinement for anomaly detection in medical images

Area of Science:

• Medical Imaging Analysis
• Machine Learning for Healthcare
• Computer-Aided Diagnosis

Background:

• Medical anomaly detection is vital for diagnosis but hindered by costly annotations of abnormal images.
• Current methods often ignore unlabeled anomalous images during training, limiting performance.
• High-cost annotations restrict the use of available anomalous data in training.

Purpose of the Study:

• To develop a novel anomaly detection method that leverages both normal and unlabeled anomalous medical images.
• To introduce a one-class semi-supervised learning framework for enhanced medical anomaly detection.
• To improve the performance of anomaly detection by utilizing readily available unlabeled data.

Main Methods:

• Proposed Dual-distribution Discrepancy for Anomaly Detection (DDAD) using one-class semi-supervised learning (OC-SSL).
• Employed ensembles of reconstruction networks to model normal (normative distribution module - NDM) and normal+unlabeled (unknown distribution module - UDM) image distributions.
• Introduced intra-discrepancy of NDM and inter-discrepancy between NDM and UDM as anomaly scores, refined by self-supervised learning.

Main Results:

• DDAD achieved significant performance gains across five diverse medical datasets (chest X-rays, brain MRIs, retinal fundus images).
• The method demonstrated superior performance compared to a wide range of existing anomaly detection techniques.
• Outperformed state-of-the-art methods in medical anomaly detection tasks.

Conclusions:

• The proposed OC-SSL framework and DDAD method effectively utilize unlabeled anomalous images, overcoming limitations of traditional approaches.
• DDAD offers a significant advancement in medical anomaly detection, improving diagnostic assistance.
• The study provides organized benchmarks and code for reproducible research in medical anomaly detection.