Function r2primenet 

pub fn r2primenet(
    r2star: &[f64],
    mask: &[u8],
    grid: &Grid,
    b0_tesla: f64,
    model_onnx: &[u8],
    norm: &R2PrimeNetNorm,
    params: &R2PrimeNetParams,
    progress: impl FnMut(usize, usize),
) -> Result<Vec<f64>, OnnxError>

Run R2PRIMEnet: predict R2′ (Hz) from an R2* map (Hz).

Arguments

• r2star — R2* map in Hz, column-major (nx,ny,nz).
• mask — binary brain mask (same layout).
• grid — volume dimensions and voxel sizes.
• b0_tesla — field strength the R2* map was acquired at, in Tesla. An argument rather than a R2PrimeNetParams field precisely because it must not have a default: a 3 T that nobody chose is indistinguishable from a 3 T that nobody checked, which is the whole of the failure this guards. See the module docs.
• model_onnx — bytes of the exported r2primenet.onnx (1→1 chan, dynamic spatial axes).
• norm — training normalization constants.
• params — patch size and the field-strength override; see R2PrimeNetParams.
• progress — progress callback (patches_done, patches_total).

Returns

R2′ in Hz, clipped at zero and restricted to mask, in the same layout.

Deviation: the authors’ test.py (χ-sepnet-R2*) passes R2PRIMEnet’s output on to χ-sepnet without truncating it. Here negative R2′ is set to 0, since a negative reversible decay rate has no physical meaning; every non-negative value is the network’s unchanged. QSM.rs clamps R2′ the same way wherever it makes it (this and r2prime), so every consumer, χ-sepnet included, gets the clamped map. The effect on χ-sepnet is negligible: on two travelling-head scans (Philips and GE, 3 T) 0.3–0.4% of brain voxels were clamped (all above −1 Hz), χ+ and χ− changed by at most 1.5 ppb (median < 0.001 ppb), and mean χ+/χ− in deep grey matter by under 0.001 ppb, about 1e-5 of its value.

Errors

OnnxError::Domain when b0_tesla is more than B0_TOLERANCE_T from TRAINED_B0_T and R2PrimeNetParams::ignore_field_mismatch is not set — checked before anything is allocated or loaded.

Panics

If a patch dimension is not a multiple of 16 (the network’s four pooling levels).