TMAN: A Triple Morphological Feature Attention Network for Fine-Grained Classification of Breast Ultrasound Images

Accurately diagnosing various types of breast lesions is critical for assessing breast cancer risk and predicting patient outcomes, which necessitates a fine-grained classification approach. While convolutional neural networks (CNNs) are predominantly employed in fine-grained classification tasks for breast lesions, they often struggle to effectively capture and model the intricate relationships between local and global features, an aspect that is vital for achieving high classification accuracy. Additionally, Color Doppler Flow Imaging (CDFI) and Strain Elastography (SE) are two important ultrasound imaging techniques widely used in the diagnosis of breast lesions. However, their specific contributions to fine-grained classification have not been thoroughly investigated. In this paper, we introduce a Triple Morphological Feature Attention Network (TMAN) designed to enhance fine-grained classification of breast ultrasound images. The TMAN architecture comprises three key modules: Local Margin Attention (LMA), Structured Texture Attention (STA), and Fusion Attention (FA), each focused on extracting distinct morphological features. TMAN achieved an average accuracy of 74.40%, precision of 73.18%, and specificity of 96.02%, surpassing state-of-the-art methods. The findings reveal that incorporating CDFI significantly improved classification for malignant subtypes with a 10% accuracy boost, while SE had a negligible impact. These findings highlight the effectiveness of TMAN in extracting nuanced morphological features and advancing precision in breast ultrasound diagnosis. The source code is accessible at https://github.com/windywindyw/TMAN .