amdenterpriseai/aim-instinct-model-monai-wholebody-ct-segmentation repository overview

⁠Whole Body CT Segmentation

A SegResNet-based volumetric segmentation model that segments 104 distinct anatomical structures (organs, bones, muscles, vessels) from whole-body CT scans. Trained on the TotalSegmentator dataset (Wasserthal et al. 2022), the bundle runs 96x96x96 sliding-window inference with gaussian patch fusion and an RAS-oriented intensity-normalized preprocessing chain mirroring inference.json. Two variants are bundled: a 1.5 mm high-resolution model (default, ~28.7 GB GPU at full-volume inference) and a 3.0 mm low-resolution model that fits tighter GPU budgets; the displayable_configs#highres flag selects between them. Bundle published mean Dice = 0.8 (per-class mean over the 104 foreground structures on its TotalSegmentator validation split). Intended use per the bundle: an example, not for diagnostic purposes.