Module autoqsm 

AutoQSM single-step deep-learning reconstruction (onnx feature).

AutoQSM (Wei 2019) is a patch V-Net that maps a total field (ppm) directly to susceptibility (ppm) — no brain extraction, no separate background removal. The network takes a fixed 64³ input patch and returns the central 32³. Whole volumes are reconstructed by overlap-tiled sliding-window inference: 32³ output patches at stride 24 (8-voxel overlap), each from a 64³ input patch with a 16-voxel context margin, edge-padded and linearly blended across the overlap. This mirrors the authors’ util.data_predict / patch_process.

Upstream weights are Keras; QSM.rs ships a clean PyTorch re-export (scripts/onnx-export/export_autoqsm.py) so tract runs it. The network is all Conv3D+ReLU with no normalization, so no input scaling is applied.

Field sign. AutoQSM was trained on fields of the opposite sign to this crate’s (field = D * chi with D = 1/3 - kz^2/k^2, so a paramagnetic source raises the field along B0). On the authors’ own test_data/0.mat input the network returns bright deep grey matter, while the same field read in this crate’s convention (SHARP + TKD) anticorrelates with that output (r = -0.42); fed this crate’s field unchanged it returns -chi-like maps (r = -0.50 on the test phantom). The field is therefore negated on the way in: the network sees its own training convention and its output is already susceptibility in ppm with the usual sign. The network is non-linear, so negating the output instead would not be equivalent.

Weights are not bundled; the caller passes the exported autoqsm.onnx bytes.

Reference pipeline

From Wei et al., NeuroImage 2019 (doi:10.1016/j.neuroimage.2019.116064), Methods:

• Training data: 3T GE MR750, 0.86×0.86×1 mm (reconstructed), 256×256 matrix, 8 echoes.
• Input: Laplacian-unwrapped total phase of every echo, combined as ψ = Σφᵢ / (γ·B0·ΣTEᵢ) (a TE-weighted mean of φ/TE) — a total field in ppm over the whole FOV, no brain mask. Labels were STAR-QSM after SMV background removal, masked with FSL BET.
• Orientation: training inputs are axial; sagittal/coronal data “should be resliced to the axial orientation”.
• Resampling: the authors’ code/test.py and util.data_predict do not resample.

Agreement with the original

tests/models_onnx/field.rs runs it on a 36×32×28 box of in-vivo data (1 mm, 2.89 T, across the brain’s edge) against its original’s verified QSM-CI entry, AutoQSM b59f5e7 (autoqsm) (QSMxT/QSM-CI#316; the fixture is regenerated in Actions by tests/data/dl_field_qsmci_small.py): NRMSE 1.8e-4 %, max |Δ| 2.5e-7 ppm (measured 2026-10-30): float32 rounding.

Pipeline rules

The facts above that a configuration can depart from are published as rules in crate::pipeline::rules::RULES (each with its strength, basis, reason and source): autoqsm.unwrap-laplacian, autoqsm.echo-combination, autoqsm.echo-weights.

Constants

IN 🔒

MARGIN 🔒

OUT 🔒

OVERLAP 🔒

SHIFT 🔒

Functions

autoqsm

Run AutoQSM on a total field map.

blend 🔒

Blend the leading OVERLAP slabs of patch with the trailing slabs of a previously-placed neighbor along dir (0=x,1=y,2=z), ramping neighbor→patch (matches util.patch_process).

output_extent 🔒

The output-grid extent for an axis of d voxels: the smallest whole number of output patches (OUT, then one more per SHIFT) that covers it. An axis shorter than one patch gets one patch; the samples past d are edge-padded and cropped away.

ri 🔒