Gradient-informed structural prior that dynamically adjusts kernel geometry per-voxel for high b-value diffusion MRI in oncology. Leverages local tissue anisotropy to preserve sharp tumor boundaries whilst aggressively suppressing noise. 18% SNR gain with 96% tissue boundary preservation vs. fixed-size MPPCA baselines on HCP data at b-values of 5000–10000 s/mm².

Novel GNN integrating ARMA spectral filtering with reconstruction-driven regularisation — first approach to jointly address over-smoothing and over-squashing in sparse-label neuroimaging graphs. Represents entire cohorts as connected graphs where nodes encode neuroanatomical features and edges capture inter-subject similarity. 98.3% accuracy CN vs. MCI, 99.7% CN vs. AD on ADNI dataset (>3% above GCN/GAT baselines).

Swin-Transformer architecture for quantitative diffusion tensor imaging maintaining diagnostic accuracy with 50%+ fewer gradient directions (5–21 vs. standard 41). Voxel-wise TBSS analysis maps disease-specific white matter degradation for tract-level FTD classification without full acquisition. Custom loss function enforces positive-definiteness and biophysical coherence in diffusion tensor estimates.

Swin-Transformer framework for estimating FA, AxD, and MD diffusion tensor measurements using only 5 and 21 gradient directions vs. the standard 41. Outperforms linear least squares and Transformer-DTI baselines while reducing scanning time by over half. Validated on the ADNI dataset with TBSS spatial statistics analysis of white matter tracts.