PAFNet: Physics-Aware Free-Water Elimination from Single-Shell Diffusion MRI

Deepa Pandey, Soma Samanta, Ranjeet Ranjan Jha, Sudhir Kumar Pathak, B.V.Rathish Kumar, Durgesh Kumar Dwivedi

Keywords: Diffusion MRI, Free-water elimination, Physics-aware neural networks, Self-attention, Advection-diffusion modeling

Abstract: Free-water elimination (FWE) plays a vital role in diffusion MRI by reducing partial volume effects (PVEs), especially in regions with white matter and CSF overlap, thereby enabling more accurate estimation of brain microstructure and connectivity. FWE becomes even more important in regions affected by edema, where excess extracellular water is present. Although two-compartment models offer accurate FWE, they require multi-shell diffusion data, which is often infeasible in clinical settings due to larger acquisition time and complex setup. In this work, we propose a physics-aware framework for performing FWE directly from single-shell diffusion MRI. The proposed architecture combines self-attention with an advection-diffusion solver, enhanced by depthwise-separable convolutions and squeeze-and-excitation(SE) modules for refined features learning. A hybrid loss function integrating RMSE, and edge-aware Sobel terms improves accuracy and generalization. Trained on HCP and evaluated on HCP, the method surpasses existing model and learning-based FWE approaches, delivering higher accuracy and more reliable tractography.

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Interpretability and Explainable AI

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 141