Module inversion 

Dipole inversion stage

Dispatcher for standard dipole inversion algorithms. Also contains TGV and QSMART pipeline runners which combine multiple stages internally.

MEDI Hz↔radians conversion is handled internally — the caller passes ppm fields and receives ppm results.

Constants

IQSM_ERODED_RAD 🔒

IQSM_PHASE_SIGN 🔒

The authors’ fixed inference conventions for the iQSM/iQFM LoT-Unet family: phase sign -1 and a 3-voxel mask-erosion radius. These are training-time constants, not user knobs, so the pipeline runners bake them in.

Functions

run_autoqsm 🔒

Source the AutoQSM weights and run inference (requires the onnx feature). field is the total field (AutoQSM does its own background removal).

run_dipole_inversion

Run standard dipole inversion on a local field.

run_iqfm

Run iQFM: joint unwrapping + background removal from wrapped phase in one network (the tissue-field head of the iQSM LoT-Unet). Returns the local (background-removed) field in ppm — feed it to any dipole inversion. This is a deep-learning replacement for the unwrap + BFR stages, not an inversion. Requires the onnx feature + iqfm weights.

run_iqsm

Run iQSM single-step reconstruction from wrapped phase (joint unwrap + background removal + dipole inversion). Multi-echo inputs are reconstructed per echo and magnitude·TE²-combined. Requires the onnx feature + iqsm weights.

run_iqsm_plus

Run iQSM+ single-step reconstruction from wrapped phase (orientation-adaptive variant; the B0 direction is a genuine network input). See run_iqsm.

run_ir2qsm 🔒

Source the IR2QSM weights and run inference (requires the onnx feature).

run_lpcnn 🔒

Source the LPCNN proximal-CNN weights and run inference (requires the onnx feature). LPCNN’s k-space data-consistency step uses the B0 direction.

run_modl_qsm 🔒

Source the MoDL-QSM prior-CNN weights and run inference (requires the onnx feature). Output is the STI χ33 component.

run_nextqsm

Run NeXtQSM end-to-end from the total field (it does its own background removal), sourcing both weight files from the registry. Unlike the entries in InversionAlgorithm, NeXtQSM spans BFR + dipole inversion, so it is exposed as a standalone reconstruction rather than a dipole-inversion-stage option.

run_qsmgan 🔒

Source the QSMGAN generator weights and run inference (requires the onnx feature).

run_qsmnet 🔒

Source a QSMnet-family model’s weights and run inference (requires the onnx feature). model_id selects the registry entry (qsmnet / qsmnet-plus), norm supplies that checkpoint’s normalization constants.

run_tgv

Run TGV single-step QSM reconstruction.

run_xqsm 🔒

Source the xQSM weights and run inference. Requires the onnx feature; weights come from the model registry (local $QSM_MODEL_DIR/cache, or the download feature).