Spatio-Temporal Ultrasound Segmentation: From Flow-Based Tracking to Promptable Transformers and State-Space Models

Ovarian cancer is one of the leading causes of cancer-related mortality and an early diagnosis is highly dependent on accurate characterization of adnexal masses through ultrasounds. Current automated segmentation approaches rely on single frame analysis, failing to exploit the temporal dynamics inherent in ultrasound video data and suffering from challenges including speckle noise, probe motion and diverse lesion characteristics. This thesis investigates three strategies to incorporate temporal coherence into ultrasound segmentation. First, we implement a post-processing tracking technique using SEA-RAFT optical flow estimation to warp and fuse U-Net masks across frames, reducing temporal inconsistency. The second method implemented is an extension of the promptable SAM2 model with a self-sorting memory bank (Med-SAM2) to leverage the most informative historical frames. The last implemented method is a state-space-based model embedding state-space modules and boundary-aware losses into a Video Object Segmentation network (ViViM) for efficient long-range temporal modeling. A comprehensive evaluation of the segmentation results is performed and we demonstrate that integrating temporal information enhances segmentation accuracy and consistency compared to frame-only methods. These contributions advance automated ovarian cancer diagnosis by providing more robust, temporally coherent segmentation models that could enhance diagnostic consistency and efficiency while supporting medical professionals with reliable automated analysis tools.