A Structure-Aware Relation Network for Thoracic Diseases Detection and Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate instance-level detection and segmentation of thoracic diseases in chest X-rays are vital for automated diagnosis.
• Existing methods may not fully leverage the inherent structural and relational information within medical images.

Purpose of the Study:

• To propose a novel Structure-Aware Relation Network (SAR-Net) that enhances Mask R-CNN for improved thoracic disease detection and segmentation.
• To introduce ChestX-Det, a chest X-ray dataset with instance-level annotations for training and evaluating diagnostic models.

Main Methods:

• Developed SAR-Net, incorporating anatomical structure, contextual, and disease relation modules to encode spatial and co-occurrence relationships.
• Created ChestX-Det, a dataset with ~3500 images and instance-level annotations for 13 common diseases, derived from NIH ChestX-ray14.
• Evaluated SAR-Net on ChestX-Det and DR-Private datasets.

Main Results:

• SAR-Net demonstrated significant performance improvements over the Mask R-CNN baseline.
• The proposed relation modules effectively capture and utilize disease and anatomical relationships for better segmentation and detection.
• Experimental validation confirmed the efficacy of SAR-Net in enhancing diagnostic accuracy.

Conclusions:

• SAR-Net offers a robust approach for instance-level thoracic disease detection and segmentation in chest X-rays.
• The ChestX-Det dataset provides a valuable resource for advancing research in automated medical image diagnosis.
• The integration of domain knowledge through relation modules is a promising direction for medical AI systems.