QSM-CI method: INR-QSM (v1)

M. Zhang, Feng R., Li Z., Feng J., Wu Q, Zhang Z., C W, J. Wu, Yan F., Liu C., Zhang Y., Wei H.

INR-QSM — a subject-specific UNSUPERVISED deep-learning dipole inversion using an implicit neural representation. No pretrained weights: a sine-activated coordinate MLP (SIREN) is OPTIMIZED per-subject so that the susceptibility it represents, pushed through the QSM dipole forward model, reproduces the input local field, with edge-weighted TV and gradient-domain regularizers. Consumes the local (tissue) field in ppm and produces susceptibility (ppm). The dipole kernel is built from the B0 direction and voxel size. Runs CPU-only by default (the reference is GPU-oriented, ~10 GB VRAM; see RUNTIME/GPU CAVEAT in README). Because it optimizes per input volume with a patch-based non-local phase-compensation scheme, runtime is long. QSM-CI reconstruction method inr-qsm. Browse and run it at https://qsmxt.github.io/QSM-CI/submission.html?method=inr-qsm.