Joint reconstruction of multi-contrast relaxation-diffusion imaging

Abstract

Diffusion MRI across multiple gradient directions, b-values and echo times (TEs) provides complementary tissue contrast but requires extensive sampling. Joint k-space and q-space undersampling requires reconstruction that preserves diffusion- and TE-dependent signal variation while exploiting shared anatomical information. Here we introduce DIFFUSE (Diffusion-relaxometry Implicit Field Fitting from Undersampled k-q Space across Echoes), a subject-specific joint k-q-TE image reconstruction framework. It fits a single implicit neural representation (INR) directly to complementary undersampled multi-coil k-space measurements across diffusion encodings and echo times. The INR (a self-supervised neural network) maps spatial location, b-value and TE to baseline b0 signals and spherical harmonic coefficients for directional diffusion-weighted image reconstruction. The measurement-specific forward model incorporates phase, coil sensitivities and motion and eddy current related transforms. We evaluated DIFFUSE using retrospective Human Connectome Project data, controlled motion simulations and in vivo multi-shell and multi-TE acquisitions. At sixfold in-plane acceleration, DIFFUSE reduced normalized mean squared error by 57-73% relative to SENSE across two diffusion shells in the retrospective experiment. In the controlled motion experiment, motion-aware DIFFUSE reduced normalized mean squared error by 66-72% relative to static DIFFUSE. The prospective study demonstrated multi-shell reconstruction at 1.3 mm isotropic resolution with significantly high signal-to-noise ratio (SNR). In a separate multi-TE experiment, a single representation jointly reconstructed all nine contrasts corresponding to 3 b-value/TE combinations. DIFFUSE thus provides a common reconstruction strategy for accelerated protocols spanning multiple diffusion encodings and echo times.