UNesT

Transformer-based whole brain, renal, and multi-organ segmentation

View repository ↗

MIT License

UNesT is the updated model trained on the same data as SLANT; the SLANT page links here as the successor.

UNesT (local spatial representation learning with a hierarchical Transformer) is a segmentation model that achieves state-of-the-art results on whole brain segmentation (132 regions plus TICV/PFV), renal substructure segmentation, and multi-organ (BTCV) segmentation. A pre-built Singularity container is provided for whole brain segmentation, and the renal and multi-organ variants are also available as MONAI Bundles.

Usage

Whole brain segmentation runs from a pre-built Singularity image (download the .sif from the repository):

singularity run -e --contain \
    --home /path/to/inputs/directory/ \
    -B /path/to/inputs/directory/:/INPUTS \
    -B /path/to/working/directory/:/WORKING_DIR \
    -B /path/to/output/directory/:/OUTPUTS \
    -B /tmp:/tmp \
    --nv \
    /path/to/wholebrain.sif \
    --ticv --w_skull --overlap 0.5 --device 1

For renal substructure and multi-organ segmentation, use the inference scripts in the repository or the linked MONAI Bundles.

Xin Yu, Qi Yang, Yinchi Zhou, Leon Y. Cai, Riqiang Gao, Ho Hin Lee, Thomas Li, Shunxing Bao, Zhoubing Xu, Thomas A. Lasko, et al. “UNesT: local spatial representation learning with hierarchical transformer for efficient medical segmentation” Medical Image Analysis, 2023.