Breast Tumor Classification Using Dynamic Ultrasound Sequence Pooling and Deep Transformer Features

Breast ultrasound (BUS) imaging is widely utilized for detecting breast cancer, one of the most life-threatening cancers affecting women. Computer-aided diagnosis (CAD) systems can assist radiologists in diagnosing breast cancer; however, the performance of these systems can be degrade by speckle noise, artifacts, and low contrast in BUS images. In this paper, we propose a novel method for breast tumor classification based on the dynamic pooling of BUS sequences. Specifically, we introduce a weighted dynamic pooling approach that models the temporal evolution of breast tissues in BUS sequences, thereby reducing the impact of noise and artifacts. The dynamic pooling weights are determined using image quality metrics such as blurriness and brightness. The pooled BUS sequence is then input into an efficient hybrid vision transformer-CNN network, which is trained to classify breast tumors as benign or malignant. Extensive experiments and comparisons on BUS sequences demonstrate the effectiveness of the proposed method, achieving an accuracy of 93.78%, and outperforming existing methods. The proposed method has the potential to enhance breast cancer diagnosis and contribute to lowering the mortality rate.