Fig. 1: Overview of the proposed DL model presented in this study.

a Training stage: we used 3D patches randomly sampled from the digital subtraction bone-removal CTA scans to train the network. b Inference stage: uniform-stride sampling was used and then the prediction of those samples was merged to obtain the final prediction of the whole volume. c Illustration of the architecture of the end-to-end aneurysm prediction model. The proposed segmentation model had a similar encoder–decoder architecture as U-Net³⁰, and residual blocks²⁸ and a dual attention block²⁹ were used to improve the performance of the network. IA intracranial aneurysm, SMO spatial matrix operation, CMO channel matrix operation.