QSM-CI method: DIP-UP (PyTorch) (v2)
Year: 2026
Type: software
Abstract: DIP-UP — deep-learning PHASE UNWRAPPING packaged as a field-mapping submission. A PRETRAINED 3-D CNN predicts the per-voxel integer wrap-count of a single-echo wrapped phase (two base variants: PHU-NET3D = 2 input channels [wrapped phase + its Laplacian], the paper default; PhaseNet3D = 1 channel [wrapped phase]); a TEST-TIME Deep-Image-Prior (DIP) loop then fine-tunes that network on the specific input under a Laplacian-consistency loss and a masked total-variation loss. DIP-UP only unwraps ONE echo, so this wrapper runs it on EACH echo and then does the standard per-voxel linear fit of unwrapped phase over TE, normalized by B0, to produce the total field in ppm — the SAME echo-fit + ppm math as the reference laplacian-fieldmap / romeo-fieldmap submissions. The novelty benchmarked is DIP-UP's unwrap operator swapped in for the Laplacian/ROMEO unwrap; the downstream echo-fit is identical. GPU-capable, CPU-optional (the reference recommends a 16–24 GB GPU; see README RUNTIME/GPU CAVEAT). DOMAIN-SHIFT CAVEAT: the base net was trained on real brain GRE phase at a specific resolution/TE/field, so its wrap-count prediction may not transfer to the sim phantom's wrap patterns — see README. QSM-CI reconstruction method dip-up. Browse and run it at https://qsmxt.github.io/QSM-CI/submission.html?method=dip-up.
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7204162210