Unsupervised anomaly detection for posteroanterior chest X-rays using multiresolution patch-based self-supervised...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Increasing data volumes in medical imaging necessitate automated screening systems.
• Radiologists require tools to efficiently focus on diagnosing abnormal findings.
• Anomaly detection is crucial for identifying unexpected patterns in medical data.

Purpose of the Study:

• To propose an unsupervised anomaly detection method for posteroanterior chest X-rays (CXR).
• To leverage multiresolution patch-based self-supervised learning for robust feature extraction.
• To address the challenge of detecting diverse anomalies with unknown shapes and scales.

Main Methods:

• Utilized multiresolution patch images of varying sizes for training and testing.
• Applied self-supervised contrastive learning to learn generalized and robust patch features.
• Evaluated the method on a public dataset of posteroanterior CXR images.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art anomaly detection techniques.
• Achieved consistent and strong overall performance across different evaluation criteria.
• Effectiveness of multiresolution patch-based features was validated for anomaly detection.

Conclusions:

• Multiresolution patch-based self-supervised learning is effective for anomaly detection in CXR images.
• The method offers a robust solution for identifying diverse anomalies in medical imaging.
• This approach can enhance automated screening systems, aiding clinical practice.