TransUNet for automated segmentation of intracranial aneurysms in digital subtraction angiography: A dual-center validation study.

PURPOSE: This study systematically evaluated five deep learning models to validate their consistency in deriving clinical morphological and quantitative digital subtraction angiography (QDSA) hemodynamic biomarkers.

METHOD: In this dual-center retrospective study (2,777 images; 1,539 patients), five models-UNet, VNet, DeepLabV3, SwinUNet, and TransUNet-were trained on an internal cohort (n = 1,212; split 6:4) and validated on an external cohort (n = 327). Beyond Dice coefficients, a comprehensive framework assessed algorithmic robustness and clinical reliability in deriving morphology and QDSA parameters-specifically cerebral blood flow, cerebral blood volume, mean transit time, and time-to-peak-using the Intraclass Correlation Coefficient (ICC).

RESULTS: TransUNet achieved the highest Dice scores internally (0.839) and externally (0.878), significantly outperforming other models (all P < 0.001). TransUNet delivered consistent performance across imaging systems and remained the top-performing model in preoperative, postoperative, and coil-treated scenarios. Regarding morphological assessment, TransUNet yielded high agreement for height, width, and max dome diameter (ICC > 0.95), but it uniquely achieved high reliability for neck width (ICC > 0.82 externally), whereas other models exhibited instability. Furthermore, TransUNet achieved superior hemodynamic consistency, with ICCs ranging from 0.774 to 0.961 across all preoperative and postoperative hemodynamic parameters.

CONCLUSION: TransUNet establishes a new benchmark for automated angiographic analysis by combining local precision with global context. Its capability to generate reliable functional and geometric biomarkers suggests high potential for real-time intraoperative guidance.