Fig. 1: Data flow diagram showing our approach to achieve bone and vessel segmentations.

The training process was divided into two parts with two separate training cohorts. Two models derived from ResU-net were accessed for performance evaluation by fivefold cross-validation and were subsequently merged to form the first layer of CerebralDoc. The validation of bone and vessel segmentation which against to the annotated source images was used to make sure the algorithm indicators acceptable. We deployed the final two models to an independent cohort that contained 1826 cases, including almost all major CTA platforms on the market, to manifest the credibility of CerebralDoc by accuracy analysis.