Enum SynthSegInference 

pub enum SynthSegInference {
    FullVolume,
    SlidingWindow {
        patch: usize,
        overlap: usize,
    },
}

How the network is applied to the preprocessed (1 mm, RAS, padded) volume.

Variants

FullVolume

One network pass over the whole padded volume, as the SynthSeg CLI runs it; the output matches FreeSurfer’s mri_synthseg (every label Dice ≥ 0.9999). Peak memory grows with the volume (about 0.95 KB per padded voxel): 6–12 GB on 1 mm head scans, so it does not fit a 4 GB WASM heap even with the head crop and half-precision posteriors.

SlidingWindow

Gaussian-weighted sliding window: the network runs on cubes of patch voxels (at least 64, rounded up to a multiple of 16) that overlap their neighbours by at least overlap voxels, and each voxel’s posteriors are the Gaussian-weighted (σ = patch/8) mean over the windows that cover it. Peak memory for the network follows the patch. An axis the volume does not exceed is covered by one window; a volume one window covers runs as FullVolume does.

Window origins are on the network’s max-pooling lattice (multiples of 16 voxels from the whole-volume input’s origin), so every window pools the same voxels together that the whole-volume pass does; the step between windows is a multiple of 16 no larger than patch − overlap, spread evenly over the axis. Against windows at a step of half a patch this gave no measurable gain on three 1 mm scans: at 128³ those steps were already on the lattice, and at 96³ the lowest label rose on one scan (Dice 0.948 → 0.963) and fell on another (0.937 → 0.928). The error at 96³ comes from the window’s limited context, not from pooling phase. The lattice is kept because it costs nothing and is the placement under which a window sees exactly what the whole-volume pass sees.

The overlap is a minimum in voxels, so the window count is explicit. At 96³, an overlap of 64 instead of 48 raised mean Dice by up to 0.003 on two of the three scans and cut the worst label’s volume error (9.8 → 6.1 %), but needs 1.5–2× the windows and about 1.6× the time. At 128³ the two give the same windows on 192- and 256-voxel axes.

Fields

patch: usize

Edge of the cubic patch, voxels at 1 mm.

overlap: usize

Minimum overlap between neighbouring windows, voxels (less than patch).

Implementations

impl SynthSegInference

pub const LOW_MEMORY: Self

The memory-bounded window, for a 32-bit WASM heap (see SynthSegParams::low_memory): 96³ patches overlapping by at least 48. Against FreeSurfer’s mri_synthseg (which the whole volume matches) on three 1 mm scans: mean Dice over the labels 0.984–0.988, lowest label 0.93–0.96, single labels’ volumes off by up to 6–10 % (the inferior lateral ventricles and the cerebellum). 128³ windows reach 0.993–0.996 (lowest 0.98, volumes within 3 %) but, even with half-precision posteriors and the head crop, took a 184×256×128 scan (padded to 192×256×128) to 3.83 GB of the 4 GB heap; a larger grid does not fit, so 128³ is an option only for grids up to about 192×256×128. Overlap 48 rather than 64: on 192- and 256-voxel axes it places as many windows as a step of half a patch, where 64 needs 1.5–2× as many (see SynthSegInference::SlidingWindow).

Trait Implementations

impl Clone for SynthSegInference

fn clone(&self) -> SynthSegInference

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for SynthSegInference

impl Debug for SynthSegInference

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

Formats the value using the given formatter.

impl PartialEq for SynthSegInference

fn eq(&self, other: &SynthSegInference) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl Serialize for SynthSegInference

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

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for SynthSegInference