Segmentation of bAVMs in 3D rotational angiography remains an open problem

Establish a reliable computational method that accurately segments brain arteriovenous malformations in 3D rotational angiography images, enabling robust identification of feeding arteries, draining veins, and the nidus for clinical analysis and planning.

Background

Brain arteriovenous malformations present complex vascular structures, including feeding arteries, draining veins, and a nidus, whose accurate delineation is critical for morphology- and topology-based clinical assessments and treatment planning. Intraoperative 3D rotational angiography offers high-resolution visualization of these structures but requires precise segmentation to extract clinically relevant measurements.

Existing non-deep-learning methods such as thresholding and region growing struggle in bAVM cases due to inconsistent contrast distribution, noise, and the intricate, entangled morphology of the nidus and venous drainage. Moreover, bAVMs are rare and their irregular hemodynamics complicate manual annotation, resulting in limited, labor-intensive, and potentially inconsistent ground truth data, which has hindered the development of robust automated methods. As a result, segmentation of bAVMs in 3DRA remains unresolved and is of high clinical relevance.