FastSurfer Phase 4a: cerebellum 28-class sub-seg (axial + coronal +
sagittal FastSurferCNN, 3-conv blocks with 5×5 + 1×1 kernels, 64
filters). Runs on a pre-cropped 128³ cerebellum ROI from brainix. All
compute in C/WGSL/SIMD128; orient + thick-slice in WASM SIMD, forward +
view_accumulate + argmax in WebGPU. Verified bit-identical (100.0000%)
against Python on the brainix ROI.
✓ Ready. Pre-cropped 128³ cereb ROI (2 MB) ships as demo input.
GPU acc buffer = 128³ × 28 × 4 = 7.3 MB (single WebGPU storage buffer,
no chunking).
Why no "upload your T1" button here: CerebNet is a downstream
sub-segmentation of FastSurfer. Locating the cerebellum requires
FastSurfer's aparc labels (IDs 7, 8, 46, 47) to compute a bounding box
on the conformed volume. Without an aparc input, raw T1 cannot be
cropped to the 128³ cereb ROI this demo expects. For the full workflow,
run the FastSurfer demo
first — its output is what feeds this model in Python. The demo here
exercises the CerebNet architecture on its expected input, validated
bit-exact against PyTorch.