Anatomy-Aware Deep Learning Model (Anatomy XNet) for Thoracic Disease Detection in Chest X rays
DOI:
https://doi.org/10.52783/rev-alap.150Keywords:
Deep Learning, Chest X-ray, Thoracic Disease, Anatomy-Aware Model, Convolutional Neural Network (CNN), Attention Mechanism, Medical Imaging, Disease Classification, Artificial Intelligence in Radiology, Interpretability in AIAbstract
Deep learning has shown great promise in the automatic detection of thoracic diseases using chest X-rays (CXR). However, most existing models lack anatomical awareness, often treating the entire image uniformly without accounting for organ-specific localization, which can lead to false positives and reduced interpretability. This paper presents Anatomy-XNet, a novel anatomy-aware convolutional neural network (CNN) that integrates anatomical priors to enhance thoracic disease classification. The model incorporates an Anatomy-Aware Attention Module (A³M) and Probabilistic Weighted Average Pooling (PWAP) to focus on key anatomical structures such as the lungs, heart, and diaphragm, using organ-level annotations. We evaluated Anatomy-XNet on three major public datasets: NIH ChestX-ray14, CheXpert, and MIMIC-CXR, achieving AUC scores of 85.78%, 92.07%, and 84.04%, respectively. Compared to baseline DenseNet-121 models, Anatomy-XNet consistently outperformed in both classification accuracy and localization precision. The model also demonstrated improved interpretability through heatmap visualization, aligning closely with expert annotations. These results underscore the importance of incorporating anatomical context in medical image analysis and pave the way for more accurate and clinically useful AI tools in radiology.
