Pulmonary Textures Classification via a Multi-Scale Attention Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate classification of pulmonary textures is vital for diagnosing diffuse lung diseases (DLDs) using computer-aided diagnosis (CAD) systems.
• Existing deep learning models struggle to learn discriminative features for complex pulmonary textures, limiting clinical application.

Purpose of the Study:

• To develop a novel deep learning architecture, the multi-scale attention network (MSAN), for enhanced pulmonary texture classification.
• To improve the discriminative feature representation for distinguishing complex lung textures in high-resolution CT images.

Main Methods:

• Designed a multi-scale attention network (MSAN) incorporating stacked residual attention modules and a multi-scale fusion module.
• Utilized visualization techniques to enhance the transparency and interpretability of the deep learning model.
• Evaluated the MSAN architecture on a large dataset of high-resolution CT images.

Main Results:

• Achieved an average classification accuracy of 94.78% for 7 categories of pulmonary textures.
• Obtained an average F-value of 0.9475, demonstrating robust classification performance.
• Visualization results provided intuitive explanations of the network's decision-making process.

Conclusions:

• The proposed MSAN architecture effectively learns multi-scale features and optimal feature selection for precise pulmonary texture classification.
• MSAN achieves state-of-the-art performance in classifying pulmonary textures on high-resolution CT images, offering significant potential for DLD diagnosis.
• The developed visualization techniques enhance model interpretability, crucial for clinical adoption of AI in medical imaging.