Towards improved decision making of unruptured intracranial aneurysms using automated segmentation from MRA-TOF with iterative pseudo labeling

BACKGROUND AND PURPOSE: Achieving robust automated intracranial aneurysm segmentation across diverse datasets remains a challenging task. To enable accurate 3D morphologic assessment and support clinical decision-making, a deep learning-based method for intracranial vessel and aneurysm segmentation (DIVA-seg) from TOF-MRA using a pseudolabeling approach was developed and validated. MATERIALS AND METHODS: Three TOF-MRA data sets were used: 1) labeled data for training (n=57) and testing (n=14); 2) unlabeled data for pseudolabels (n=518); and 3) labeled data for external validation (n=82). An nnU-Net (model 1) was iteratively trained for creating pseudolabels for data set 2. Cases with stable segmentation performance across iterations were selected for further training. Stable cases (n=484) were combined with data set 1 to train a second nnU-Net (model 2). Performance testing on data set 1 and 3 comprised Dice similarity coefficient (DSC), 95% Hausdorff distances (HD), 3D morphologic measures, and a blinded qualitative evaluation. RESULTS: DIVA-seg achieved a mean (standard deviation) internal vessel and aneurysm DSC of 0.925 (±0.025) and 0.880 (±0.045), respectively. On the external test set the DSC were 0.899 (±0.028) and 0.861 (±0.114), respectively. Mean HD was 0.67 mm for both test sets. Bland-Altman plots showed a high agreement between 3D morphologic measures from ground truth and model segmentations; however, a proportional bias was observed for voxel volume, surface area, sphericity, and shape index. The qualitative evaluation showed no clear preference for either ground truth or model segmentation. CONCLUSIONS: The model achieved accurate and reliable segmentation of vessels and aneurysms internally and externally while also showing high agreement between 3D morphologic measures from automatic and manual segmentations, indicating its potential clinical utility.