Annotation-efficient selection and collaborative datasets for clinical translation in brain Vessel segmentation

Brain vessel segmentation is fundamental for understanding cerebrovascular diseases and planning neurosurgical treatments. Yet, current deep learning segmentation approaches face critical challenges, including extensive annotation requirements and inconsistent quality across datasets. These barriers create a significant gap between research and clinical translation in real-world medical environments. This thesis proposes a set of contributions to advance the field of neurovascular imaging. First, to address the problem of expert vessel annotation requiring multiple hours per 3D volume and demanding specialized neuroanatomical knowledge, V-DiSNet introduces a one-shot active learning framework that uses dictionary learning to identify recurring

vessel patterns, reducing annotation requirements by 70% while maintaining segmentation accuracy. Second, to tackle the dataset fragmentation

caused by different annotation protocols across institutions, Vessel-Verse provides the largest public vessel annotation dataset comprising 950 3D images with a collaborative framework that supports version control, multi-expert annotation, and automated consensus generation. Third, building on these contributions and leveraging a cross-modality adaptation vessel segmentation framework, three studies bridge the clinical

translation gap by demonstrating practical utility in real-world clinical research. These studies include automated stroke diagnosis using 902 clinical TOF-MRA images, genetic analysis using automated vessel quantification in 230 twin pairs to identify heritable vascular traits and their clinical implications, and a multi-institutional federated learning vessel segmentation deployment across diverse hospitals.

This thesis aims to advance neurovascular image analysis through the development

of annotation-efficient methodologies and collaborative infrastructure, thereby enabling the clinical translation of research algorithms into practical healthcare solutions.