Anatomy-XNet: An Anatomy Aware Convolutional Neural Network for Thoracic Disease Classification in Chest X-Rays.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Deep learning for thoracic disease detection from chest radiographs is an active research area.
• Current methods often focus on diseased regions, unlike expert radiologists who identify anatomical structures first.
• Integrating anatomical knowledge into deep learning can enhance automatic disease classification.

Purpose of the Study:

• To propose Anatomy-XNet, an anatomy-aware attention-based network for thoracic disease classification.
• To leverage semi-supervised learning with organ-level annotations for anatomy region identification.
• To improve the accuracy and generalizability of AI in diagnosing thoracic diseases.

Main Methods:

• Developed Anatomy-XNet, an attention-based network prioritizing spatial features guided by pre-identified anatomy.
• Utilized a semi-supervised approach with small-scale organ annotations to locate anatomy in large datasets.
• Employed a pre-trained DenseNet-121 backbone with Anatomy Aware Attention (A3) and Probabilistic Weighted Average Pooling modules.

Main Results:

• Achieved state-of-the-art performance on three large-scale chest X-ray datasets (NIH, Stanford CheXpert, MIMIC-CXR).
• Obtained AUC scores of 85.78%, 92.07%, and 84.04% on the respective datasets.
• Demonstrated the effectiveness of incorporating anatomical knowledge for improved thoracic disease classification.

Conclusions:

• Anatomy-XNet effectively utilizes anatomical knowledge for superior thoracic disease classification.
• The proposed framework shows strong generalizability across different large-scale chest radiograph datasets.
• Integrating anatomy-aware attention mechanisms represents a significant advancement in medical image analysis.