Fig. 2: Study workflow.

The modular pipeline comprises three sequential stages: automated hematoma segmentation, synthetic minority data augmentation, and automated hematoma classification. Stage 1: Four state-of-the-art 3D segmentation networks (U-Mamba, nnU-Net, nnFormer, and UNETR++) were benchmarked on a preliminary dataset of 1000 NCCT scans (baseline and follow-up) from 500 sICH patients randomly selected from the full training cohort of 1,103 patients. The best-performing model (U-Mamba) was then trained on all 2206 scans from the complete training cohort to generate high-quality hematoma masks. Stage 2: Synthetic minority data augmentation employed Diffusion-UKAN to generate high-fidelity synthetic HE images, yielding two augmented training sets: UKAN-Balanced (HE: NHE = 1:1) and UKAN-Semibalanced (HE: NHE = 1:2). Stage 3: Automated classification used a Vision Transformer trained on three consecutive slices centered on maximum hematoma area (Max−1, Max, Max+1). Patient-level predictions were obtained by averaging slice-level probabilities, with Grad-CAM providing visual interpretation of discriminative regions. ROI, region of interest; HE, hematoma expansion; NHE, non-hematoma expansion.