HypVINN — 3-view hypothalamus sub-segmentation (T1-only)

FastSurfer Phase 4b: hypothalamus 27-class sub-seg (axial + coronal + sagittal HypVINN networks, dual-input InputDenseBlock with m1 / m2 branches + softmax mod_weights). T1-only mode: T1 thick-slice duplicates as both modality channels, weight_factor=[1, 0] zeros the T2 branch. All compute in C/WGSL/SIMD128. Verified bit-identical (100.0000%) against PyTorch HypVINN on brainix in T1-only mode.
✓ Ready. Brainix T1 in RAS-canonical form ships as demo input (256³ u8, 16.8 MB). GPU accumulator = 256³ × 27 × 4 = 1.77 GB (single buffer, fits under the 2 GB WebGPU cap).
⚠ T1-only mode. HypVINN was trained on paired T1+T2. This demo runs the --mode t1 code path (a documented FastSurfer CLI option), but the resulting segmentation is less accurate than with a real registered T2 scan.
idle
Demo T1: brainix 0.9375 mm RAS 256³ (bit-exact parity with PyTorch). Upload path: custom_ops_conform_u8 produces RAS 0.9375 mm 256³ u8 to match the fixture's scale (Zoom2d is hard-wired to this scale; dynamic Zoom2d deferred).

Slice viewer (after run)

Log