Struct L1QsmParams 

pub struct L1QsmParams {
    pub alpha1: f64,
    pub mu1: Option<f64>,
    pub mu2: f64,
    pub mu3: f64,
    pub lambda: f64,
    pub max_iter: usize,
    pub tol_update: f64,
    pub tol_delta: f64,
    pub b0: f64,
    pub echo_times: Vec<f64>,
    pub te: EchoTime,
    pub kernel: FansiKernel,
}

L1-QSM algorithm parameters (FANSI nlL1TV.m defaults where it has them).

Fields

alpha1: f64

Gradient (TV) L1 penalty weight. nlL1TV.m gives no default, and FANSI’s example says L1-QSM needs its own tuning. The default 5e-3 was chosen from an alpha1 sweep on the EMTP travelling-head data: 2e-4 streaks, 1e-2 starts to oversmooth.

mu1: Option<f64>

Gradient-consistency ADMM weight; None is nlL1TV.m’s 100 * alpha1.

mu2: f64

Fidelity-consistency ADMM weight.

mu3: f64

L1 proximal ADMM weight.

lambda: f64

Phase-rejection strength: the data weight is lambda times the normalised magnitude. Smaller values reject more inconsistent voxels. Default 0.8, the value of FANSI’s in-vivo demo (script_DICOM_Philips.m l. 445-447: lambda = 0.8; params.weight = lambda*params.weight/max(params.weight(:))); nlL1TV.m itself has no default (“recommended = lambda*magnitude_data”, l. 18-19). QSM.rs previously used 1.0. Measured against 1.0 (ROMEO field, default V-SHARP): on the qsm-forward phantom χ XSIM 0.338 → 0.339, NRMSE 78.2 → 78.1 % (on its true local field 0.808 → 0.809, NRMSE 38.8 → 39.2 %); on an in-vivo 3T Philips scan, agreement with the UK Biobank pipeline’s χ XSIM 0.446 → 0.461, NRMSE 103.0 → 97.7 %.

max_iter: usize

Number of outer ADMM iterations.

tol_update: f64

Percent-update convergence stopping tolerance.

tol_delta: f64

Inner Newton convergence tolerance.

b0: f64

Field strength in Tesla (scan parameter, for the radian conversion)

echo_times: Vec<f64>

Echo times in seconds, in acquisition order (scan parameter, for the radian conversion).

te: EchoTime

Which echo time the radian conversion uses. Default EchoTime::First. FANSI’s multi-echo example, script_DICOM_Philips.m, scales the field by deltaTE = TE(echo2) - TE(echo1) (FANSI-toolbox 93044a3, line 364) before nlL1TV (line 450), and its comment on that line says “You may use also any individual echo time”. On the qsm-forward phantom (echoes 4/12/20/28 ms; FANSI’s discrete kernel, measured for QSM.rs#418 with the integration tests’ settings; XSIM / NRMSE / r), the magnitude weight at the first echo does best (0.776 / 38.9% / 0.923), against 0.764 / 46.9% / 0.889 at the echo spacing. (With the continuous kernel, #276: 0.782 / 37.4% / 0.928 against 0.776 / 40.2% / 0.916.)

The magnitude weight is required, although nlL1TV.m weights by ones when given none (lines 109-113): with a flat weight L1-QSM is worse on that phantom (0.767 / 42.0% / 0.912 at the first echo, 0.763 / 43.1% / 0.906 at the spacing). With the continuous kernel the flat weight’s NRMSE exceeded 100% at either echo time (0.738 / 108% / 0.636 at the first echo, #276).

kernel: FansiKernel

The dipole kernel. Default FansiKernel::Discrete, FANSI’s dipole_kernel_angulated as the original is run with; FansiKernel::Continuous is the crate’s continuous kernel (see FansiKernel for the decision and the measurements).

Trait Implementations

impl Clone for L1QsmParams

fn clone(&self) -> L1QsmParams

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for L1QsmParams

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for L1QsmParams

fn default() -> Self

Returns the “default value” for a type.

impl Serialize for L1QsmParams

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.