Facing Differences of Similarity: Intra- and Inter-Correlation Unsupervised Learning for Chest X-Ray Anomaly...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Anomaly detection in chest X-rays aids medical interpretation.
• Pre-trained networks for feature extraction face limitations due to similarity differences in detailed X-rays.

Purpose of the Study:

• To propose an intra- and inter-correlation learning framework for robust chest X-ray anomaly detection.
• To address limitations of pre-trained networks in capturing detailed X-ray features.

Main Methods:

• Introduced an Anatomical-Feature Pyramid Fusion Module for feature fusion, integrating local details and global context.
• Developed an intra-correlation learning strategy to map image features to semantic centers for lesion discovery.
• Employed inter-correlation learning to mitigate similarity differences between images from pre-trained networks.

Main Results:

• The proposed framework demonstrated superior and effective anomaly detection across various scenarios.
• Achieved improved feature representations by leveraging anatomical structures and inter-image correlations.
• Outperformed 18 state-of-the-art methods on three diverse datasets.

Conclusions:

• The intra- and inter-correlation learning framework offers a robust solution for chest X-ray anomaly detection.
• The method effectively handles differences in similarity, enhancing diagnostic accuracy in diverse disease environments.
• This approach shows significant potential for improving the interpretation of medical imaging.