Set up and run local AI segmentation

1. If no environment is installed

Use the official TotalSegmentator installation guide linked below. Create a separate virtual or Conda environment, install Python/PyTorch for your platform, then TotalSegmentator inside it. Do not install into the VoxPlane bundle.

The project-tested combination is TotalSegmentator 2.18.0 on macOS/MPS. The latest upstream version is not necessarily validated here. Windows environment discovery is automated-tested but real-device segmentation is not yet verified. Ask an administrator to prepare the environment if needed.

2. Prepare the CT fast model

Follow the official weight-download instructions and ensure the model root contains Dataset297_TotalSegmentator_total_3mm_1559subj, including its model files and configuration.

VoxPlane does not install dependencies or download models during detection or segmentation. Prepare models on a connected machine for offline use as described upstream.

3. Detect and select the environment

Open Settings → AI → Local automatic segmentation → Check environment. Use a discovered environment, or expand Advanced to select its interpreter and model directory.

On Windows, virtual environments normally use Scripts/python.exe and Conda uses python.exe at the environment root; on macOS choose bin/python. Run the check again and check dependencies, model and device; explicitly change an unavailable saved device.

4. Run segmentation

In CT MPR / 4D → Segmentation, click Automatic segmentation. Check series/phase, choose organs and start. Input is the complete original-HU volume for the phase, not the cropped or windowed screen image.

5. Review and save

After success, inspect organs, colors and visibility in the list. Review location and extent across slices and refine with Paint/Erase if needed. Save a workspace or export SEG/NRRD. A successful environment check does not prove accuracy; repeat runs do not overwrite existing segments.

Failure or missing results

On failure, inspect Details, correct environment/model/device/memory issues and retry. A GPU failure does not automatically switch to CPU.

If you switched tabs/phases, return to the original target and finish editing so results can import. Cancellation does not import partial results. Restoring a saved workspace does not require another AI run.