Projection-based quality assessment of cerebrovascular segmentations

Reliable cerebrovascular analysis from 3D time-of-flight mag netic resonance angiography depends on the anatomical completeness and consistency of vascular representations. However, automated ves sel segmentation may contain errors that hinder downstream analysis tasks. Identifying such errors directly in 3D is challenging due to the complex topology of the cerebrovascular anatomy. In this work, we pro pose a canonical multi-view framework for automated assessment of cere brovascular segmentation quality. Rather than reasoning directly over the complete 3D vascular tree, we reformulate localized quality assessment as detection in canonical two-dimensional projection views, where anatom ical structures exhibit reproducible appearance. A lightweight detection model is applied independently to complementary views, whose agree ment is used to identify potentially erroneous segmentations for manual inspection. To alleviate annotation scarcity, we propose view synthesis by perturbing projection angles and slabs around anatomy-guided ref erence views, generating anatomically valid training examples from the same registered 3D volume. We demonstrate the framework for identi fying spurious superior sagittal sinus segmentations in TOF-MRA and evaluate it on 99 subjects with manually annotated quality labels. The proposed approach achieved an F1 score of 88.31, outperforming vision language models (VLM) operating in few-shot settings or adapted with LoRA, while relying on a lightweight task-specific detector instead of large VLMs. Our results show that canonical projection views reformu late 3D cerebrovascular quality assessment into a robust and efficient 2D detection problem, facilitating scalable construction of reliable cere brovascular analysis pipelines. All code is available at github.com/erc caravel/vascular-qc.