Intracranial 4D flow MRI is widely recognized for its ability to quantify hemodynamics in ICAD patients. However, the manual process of vessel segmentation, especially in the presence of stenoses, is labor-intensive and prone to user variability. To tackle these issues, a team of researchers has developed a highly accurate, fully automated segmentation method employing deep learning techniques.
The study involved the retrospective selection of 154 dual-VENC 4D flow MRI scans, including 68 ICAD patients with stenosis and 86 healthy controls. Manual segmentations were used as the ground truth for training the deep learning model, which utilized a 3D U-Net architecture. The model’s performance was tested on 20 randomly selected cases, split equally between controls and patients. Key metrics, such as cross-sectional areas and flow parameters, were assessed in the Circle of Willis (CoW) and the sinuses, with flow conservation error also calculated. Statistical comparisons were made using Dice scores (DS), Hausdorff distance (HD), average symmetrical surface distance (ASSD), Bland-Altman analyses, and interclass correlations.
The results were impressive, with the deep learning model requiring approximately 10 hours for training and averaging a rapid 2.2 ± 1.0 seconds for automated segmentation. The segmentation performance showed no significant differences compared to manual segmentations by two independent observers. For controls, the mean DS was 0.85 ± 0.03 for the CoW and 0.86 ± 0.06 for the sinuses. For patients, the mean DS was 0.85 ± 0.04 (CoW) and 0.82 ± 0.07 (sinuses). Assessments of flow parameters revealed minimal bias and tight limits of agreement in both cohorts. In stenosed vessels, the automated segmentations showed very good agreement (ICC: 0.93) with black blood vessel wall imaging (VWI) segmentations, despite a consistent overestimation of 28.1 ± 13.9%.
The study demonstrates the successful application of deep learning for fully automated segmentation of stenosed intracranial vessels using 4D flow MRI data. The statistical analysis indicates very good agreement between the deep learning model and manual segmentations, underscoring the model’s robustness and accuracy. To enhance performance and generalization, future research will incorporate more ICAD segmentations and explore other intracranial vascular pathologies.
This significant advancement in radiological imaging is poised to streamline the assessment process for ICAD patients, providing clinicians with more reliable and efficient diagnostic tools. The full research article is available at Frontiers in Radiology.