Deep learning-based prediction of wall shear stress throughout the cardiac cycle in middle cerebral artery bifurcation aneurysms

Background and Objective: Cardiac-cycle wall shear stress (WSS) is relevant to intracranial aneurysm hemodynamic assessment, but transient computational fluid dynamics (CFD) is costly. Existing surrogates use steady-flow priors or joint steady/transient training; here we evaluate the transfer of separately pretrained spatial parameters. This study develops the Steady-to-Transient Estimation of Pulsatile Wall Shear Stress (STEP) framework, a two-stage geometric learning approach for full-cycle WSS prediction. Methods: Using the synthetic AneuG-Flow dataset, Stage 0 pretrained a geometric tokenizer, LaB-GATr backbone, and WSS decoder on 11,656 steady CFD cases. Stage 1 initialized these components from the pretrained model and fine-tuned them end to end using 449 transient training cases, with 72 validation and 57 primary test cases. Waveform embeddings multiplicatively conditioned geometric features across blocks at 80 cardiac-cycle time points. Gaussian curvature was included as an auxiliary geometry feature. Ablations evaluated steady initialization, temporal conditioning, and curvature, with rank-biserial effect sizes and Holm-adjusted p values. A fixed model was also evaluated on 231 additional cases, with registered-geometry overlap audited across stages. Results: On the primary test set, STEP achieved a whole-cycle WSS vector mean absolute error (MAE) of 1.688 Pa and root-mean-square error (RMSE) of 2.877 Pa. Time-averaged WSS, oscillatory shear index, and relative residence time MAEs were 1.040 Pa, 0.00625, and 0.06734 Pa-1, respectively. Steady initialization reduced vector MAE by 27.93% relative to training the same transient architecture from scratch; the complete blockwise conditioning configuration reduced MAE by 8.04% relative to input-only additive conditioning (rank-biserial correlations 0.995 and 0.982, respectively; Holm-adjusted p < 0.001). Curvature yielded a 3.46% reduction without a significant paired effect (adjusted p = 0.0521). The 80-time-point WSS prediction required a mean of 4.169 s (SD 0.063 s). Additional-cohort MAE was 1.712 Pa, but 185/231 geometries matched Stage 0 training, versus 1/57 in the primary test set. Conclusions: STEP demonstrated effective reuse of steady-pretrained spatial parameters for cardiac-cycle WSS prediction under the evaluated synthetic CFD protocol. The strongest controlled gains were associated with steady initialization and the complete temporal-conditioning configuration. The common inlet waveform and cross-stage geometry overlap limit generalization claims.