amdenterpriseai/aim-instinct-model-monai-vista3d repository overview

⁠MONAI VISTA-3D — Whole-body 3D CT promptable foundation segmentation

VISTA-3D is MONAI's 3D promptable segmentation foundation model. It maps 3D CT volumes to a 132-class semantic mask, supporting two inference modes: (a) automatic — label_prompt=None triggers segmentation over the bundle's 117-class everything-mode subset; (b) interactive — a non-empty label_prompt=[<class_ids>] runs targeted class-prompt segmentation, and optional points + point_labels extend that to single-object point-prompt segmentation in the original-image coordinate frame. Built on a SegResNet-like backbone with a prompt encoder; trained jointly on class prompts and point prompts across 20+ partial-label datasets. Outputs class-ID masks at the original input resolution via SaveImaged. The model weights are included in the base image and are governed by the NVIDIA OneWay Noncommercial License (§3.3 non-commercial use only); users must accept the license terms at https://huggingface.co/MONAI/vista3d⁠ and provide a Hugging Face token before use.