Part-Aware Mask-Guided Attention for Thorax Disease Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Thorax disease classification from Chest X-ray (CXR) images is crucial for computer-aided diagnosis but challenging due to complex pathologies.
• Current methods often overlook informative local regions, leading to suboptimal classification performance.
• Existing approaches primarily rely on global feature extraction from entire CXR images.

Purpose of the Study:

• To develop a novel Part-Aware Mask-Guided Attention Network (PMGAN) for enhanced thorax disease classification.
• To simultaneously learn complementary global and local feature representations from CXR images.
• To improve the accuracy and robustness of automated disease detection in thoracic imaging.

Main Methods:

• Proposed a Part-Aware Mask-Guided Attention Network (PMGAN) integrating global and local feature learning.
• Utilized soft attention modules to guide feature learning towards informative global regions.
• Implemented a mask-guided attention module, regularized by organ masks, to identify local visual cues without inference computation overhead.
• Employed a multi-task learning strategy to maximize complementary local and global representation learning.

Main Results:

• The PMGAN demonstrated superior performance in thorax disease classification compared to state-of-the-art methods.
• Experimental evaluation on the ChestX-ray14 dataset validated the effectiveness of the proposed network.
• The integration of local and global features significantly improved classification accuracy.

Conclusions:

• The novel PMGAN effectively captures both global and local visual cues for improved thorax disease classification.
• The proposed method offers a significant advancement in computer-aided diagnosis for thoracic conditions.
• PMGAN provides a robust and accurate approach for analyzing Chest X-ray images.