Repository navigation
v3.4.0
v3.4.0
Lower peak memory and faster post-processing, with identical outputs.
Highlights
- Peak memory roughly halved during parcellation (about 26 GB to 14 GB in our test).
- Faster post-processing: writing the per-level label volumes takes about 0.14 s per case instead of about 9 s.
What changed
- The three plane-wise (coronal, sagittal, axial) probability maps are now summed into a single buffer instead of allocating a full-volume array per plane. The summation order is unchanged, so outputs are bit-identical to v3.3.0.
- Level label volumes are remapped with a lookup table instead of one full-volume pass per region.
- Regional volumes are counted with a single
bincount.
Documentation
- Citation: please cite the published Human Brain Mapping article and take the citation from PubMed (link added). BibTeX is labeled.
- Added journal papers to Related Research.
- Header badge now links to PubMed; added an OpenMAP series badge.
- Section headings now have icons.
- Updated the uv install command for Windows and the PyTorch requirement (2.15+, Python 3.11+).
- All changes are mirrored in the Japanese and Chinese READMEs.
Notes
- Outputs are unchanged; verified for OpenMAP-T1 and OpenMAP-T1c on one case.
- Timings and memory figures were measured on an Apple Silicon Mac (MPS). Total runtime is dominated by network inference, so end-to-end time is largely unchanged on that setup.