Using Segment-Level Attention to Guide Breast Ultrasound Video Classification

Breast cancer is the most common cancer among women, and ultrasound serves as a critically important imaging technique for its detection and classification in dense breast tissues. Compared to individual static ultrasound images, ultrasound videos offer more comprehensive information for screening and diagnosis. Current deep learning-based methods for breast ultrasound video classification often rely on manually selected keyframes, i.e., frames with stable and clear features key to image analysis tasks. However, the selection of these keyframes depends heavily on the experience of radiologists. In this paper, we propose a breast cancer video classification method that eliminates the need for manual selection of keyframes and reduces the reliance on radiologists' experience. Specifically, we introduce a Segment Attention Generator (SAG) module to guide deep learning models to pay more attention to video segments that exhibit stable and clear appearances for predicting classification. Extensive experiments show that our proposed method boosts the performance of classifying breast ultrasound videos. The results demonstrate the effectiveness of our proposed SAG module. Additionally, we introduce a new dataset for breast ultrasound video classification to address the current scarcity of publicly available datasets. Our dataset and code are openly accessible at https://github.com/imzhangyd/SAG-Net.