A256 Automated nnU-net–based segmentation of intracranial aneurysms and cerebral vessels in 3D rotational angiography for endovascular treatment planning

Le résumé fourni par la source

Background/Introduction Accurate segmentation of intracranial vessels and aneurysms is essential for endovascular treatment planning and device selection. Current segmentation approaches are predominantly based on thresholding techniques, which frequently require manual correction and may yield inaccurate results due to inhomogeneous contrast opacification of the vasculature. These limitations may substantially affect geometric assessment and consequently influence treatment strategy. Recent advances in deep learning, particularly self-configuring nnU-Net frameworks for biomedical image segmentation, have demonstrated robust performance across a broad range of imaging applications. Objective/Purpose To develop and evaluate an nnU-Net–based framework for automated segmentation of intracranial vessels and aneurysms derived from preprocedural 3D rotational angiography (3DRA). Methods/Case Description An nnU-Net model was trained using 42 3D rotational angiography (3DRA) datasets acquired with a biplane flat-panel angiography system (Azurion, Philips Medical Systems, Amsterdam, Netherlands) from patients with intracranial aneurysms treated at our institution. Reference segmentations were generated using a semi-automated workflow combining thresholding and region-growing techniques and were independently validated by two expert raters. Segmentation performance was assessed using the Dice similarity coefficient (DSC), centerline Dice score (CLDSC), relative volume absolute error (RVAE), and average symmetric surface distance (ASSD). Results/Findings Mean segmentation time using the nnU-Net framework was approximately 50 seconds per sample. Vessel segmentation achieved a DSC of 0.89 ± 0.049, a CLDSC of 0.95 ± 0.054, and an RVAE of 0.134 ± 0.087 compared with the reference standard. Aneurysm segmentation achieved a DSC of 0.897 ± 0.054, an ASSD of 0.574 ± 0.202 mm, and an RVAE of 0.13 ± 0.113. Discussion/Conclusions/Learning Points nnU-Net–based segmentation enables fast and accurate automated segmentation of intracranial vessels and aneurysms in 3DRA. This approach may facilitate automated endovascular treatment planning workflows and improve reproducibility in neurointerventional procedures.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.