3 × 3 × 3 mm + CTNormalization)
+ symmetric pad_nd_image + sliding window + GPU unpad +
per-channel logit resample back to scanner frame + NIfTI pack.
·
→ Web Worker version (UI stays responsive)
nnUNetv2_predict CPU reference. Critical detail:
pad_nd_image pads SYMMETRICALLY (pad_below = diff//2,
pad_above = diff//2 + diff%2) — task 297 hits the pad branch
because its 3 mm canonical D axis (105) is smaller than patch D (112).
After sliding, a GPU-side unpad kernel
(sliding_unpad_accum) crops accum back to canonical before
postprocess_gpu per-channel logit resample.