Abstract

This thesis investigates data-efficient and interactive methods for medical image segmentation, addressing two major related hurdles: the scarcity of high-quality annotations for fully-supervised lesion segmentation and the labor/expertise-intensive nature of manual segmentation workflows. Our first contribution, YOLO-URSAM, is a weakly supervised intracranial hemorrhage (ICH) segmentation model that leverages the Segment-Anything Model (SAM), where we fine-tune YOLOv8 on bounding-box annotated CT scans, introduce automatic point prompting for SAM, and apply perturbation-based majority-voting for uncertainty rectification. YOLO-URSAM achieves 0.933 detection accuracy, 0.796 AUC, and a mean Dice score of 0.629, surpassing existing weakly supervised methods and popular supervised models (U-Net and Swin-UNETR) on available public data. Our second contribution, SAMIRA, is a virtual reality system with a conversational AI agent for semi-automated 3D radiological segmentation. Users issue simple voice commands to initialize masks, correct them using a human-in-the-loop approach, then view the 3D segmentations as a true-to-scale 3D mesh. In a user study, SAMIRA achieved a System Usability Scale score of 90.0±9.0, demonstrated a low cognitive load, and was praised for its intuitive guidance, educational benefits, and immersive visualization. Together, these methods combine AI foundation models, prompting, and limited human interaction to deliver accurate and efficient segmentation for clinical imaging.