NA-MIC Project Weeks

Back to Projects List

Add a MONAI Auto3DSeg inference extension to 3DSlicer

Key Investigators

• Rudolf Bumm (Cantonal Hospital of Graubünden, Switzerland)
• Andras Lasso (Queens University, Canada)
• Steve Pieper (Isomics, Inc., USA)
• Andres Diaz-Pinto (Senior Deep Learning Engineer at NVIDIA, UK)
• Umang Pandey (Universidad Carlos III de Madrid, Spain)

Presenter location: In-person

Project Description

This project aims to implement MONAI Auto3DSeg in a 3DSlicer extension. This will enable fast inference with NVIDIA GPUs and CUDA and slower inference with CPU only. Auto3DSeg is a relatively new technique in the MONAI project and our first experiments have been successful. inference is not as complicated as using the MONAOLabel inference function.
A future aim is to integrate Auto3DSeg training into the MONAILabel extension.

Objective

1. Objective A. Implement Auto3DSeg into a new 3D Slicer extension.

Approach and Plan

We have great starting code as well as 2 ready-to-use models from Andres Diaz-Pinto. We will build on that. In addition, we will train a lung lobe and airway model which should be available at the PW.

Progress and Next Steps

1. Andras developed a new extension MONAI Auto3DSeg
2. It can be downloaded via the extension manager.

3. Andres created 3 Auto3DSeg models already to enable direct inference with CT datasets

4. The best models get automatically downloaded for each process
5. They will be improved with further training

6. In future, we attempt to enable your own training of Auto3DSeg models in MONAILabel.

   

2/24/2024

Andres and Andras achieved relevant progress working on the extension during the last weeks:

The extension

• is now much faster
• has a wider range of available models
• includes low res models which use less VRAM
• some models were split into smaller pieces to be able to run them on 8 GB VRAM or CPU
• The overall quality of the models was largely improved

(using NVIDIA RTX Geforce 3070 Ti)

We´ll continue to add relevant models.

Illustrations

Algorithm Generation:

Simulate a dataset and Auto3D datalist using MONAI functions:

Background and References

https://github.com/Project-MONAI/tutorials/tree/main/auto3dseg#performance-benchmarking