DCFSeg-Net for Liver Ultrasound Sequence Segmentation in Ultrasound-Guided Surgery

Ultrasound (US)-guided liver tumor ablation surgery necessitates precise target localization for effective puncture and ablation monitoring. However, respiratory movements and intraoperative manipulations complicate this task by continuously altering the target's position and morphology. To address these challenges, we propose DCFSeg-Net, a novel network architecture designed specifically for US image sequences. This model employs a multi-task learning approach, correlating tracking and segmentation tasks to ensure accurate localization and real-time segmentation. The tracking branch utilizes a discriminative correlation filter algorithm enhanced with convolutional features for robust tracking. The segmentation branch, supported by the tracking output response map, incorporates multiple attention mechanisms and a lightweight feature enhancement module to improve feature representation. The efficacy of DCFSeg-Net was validated through both offline and online experiments. In offline inference, the network achieved a mean tracking error of 1.07mm and a segmentation accuracy Dice coefficient of 91.3%. During the online assessment, it demonstrated the capability to perform real-time tracking and segmentation at 30FPS, establishing itself as a stable and reliable method for target localization in liver US sequences.