Optimized Breast Lesion Segmentation in Ultrasound Videos Across Varied Resource-Scant Environments

Yunhao Li, Zibin Chen, Junming Yan, Ziyu Ding, Jie Li, Teng Huang, Xiaoqing Pei, Zheng Zhang, Qiong Wang, Yan Pang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 4318-4333

Medical video segmentation plays a crucial role in clinical diagnosis and therapeutic procedures by enabling dynamic tracking of breast lesions across frames in ultrasound videos, thereby improving segmentation accuracy. However, the existing methods struggle to strike a balance between segmentation accuracy and inference speed, which impedes their real-time deployment in resource-limited medical environments. To overcome these challenges, we introduce a rapid breast lesion segmentation framework named RbS. RbS employs the Stem module and RbSBlock to enhance representations through intra-frame analysis of ultrasound videos. Moreover, we have developed two versions of RbS: RbS-S boasts enhanced segmentation accuracy, while RbS-L ensures faster inference speeds. Experimental evidence indicates that RbS surpasses current leading models in both segmentation efficiency and prediction accuracy, particularly on resource-limited devices. Our contribution significantly propels the progress of developing efficient medical video segmentation frameworks suitable for various medical platforms.