FastSurfer — 3-view aparc

Whole-brain 79+19=96-class FreeSurfer-compatible segmentation. Three FastSurferVINN networks (axial / coronal / sagittal) → LIA logit accumulator → argmax → LUT remap → split_cortex_labels. WebGPU + WASM forward, JS postprocess. Verified bit-identical (99.9999 %, fp32 maxpool noise floor) against Python run_prediction.py.

✓ Bit-perfect FreeSurfer pipeline. Pre-conformed brainix orig (256³ LIA, 0.9375 mm iso, 16.8 MB) ships as the demo input — load + run reproduces the full FastSurferCNN postprocess (LUT + split_cortex_labels) without uploading anything.

⚠ Memory: the LIA logit accumulator is 1.32 GB fp32 (256³ × 79 channels). Apple Silicon UMA (16 GB+) is comfortable; on a 4 GB-tab discrete-GPU machine this might OOM. Forward dispatches stay on the GPU (~136 ms / slice on M1 Max), accumulation + postprocess on CPU.

idle

Log

Orthogonal slices (orig + segmentation overlay)

Axial (slice along S→I)

slice — / 256

Coronal (slice along A→P)

slice — / 256

Sagittal (slice along R→L)

slice — / 256