NA-MIC Project Weeks

Back to Projects List

BAMF nnUNet MRI Breast Model

Key Investigators

• Rahul Soni (BAMF Health, Singapore)
• Jithendra Kumar (BAMF Health, Singapore)
• Jeff Van Oss (BAMF Health, US)
• Gowtham Murugesan (BAMF Health, US)

Presenter location: Online

Project Description

We trained an nnUNet model to segment Breast, Fibroglandular Tissue, and Structural Tumor from MRI scans. We prepared the training set from multiple source such as Breast-MRI-NACT-Pilot, Duke-Breast-Cancer-MRI, and ISPY1-Tumor-SEG-Radiomics

Objective

1. Model Container: We would like to bring our model resources into a deployable container that can be used by research community to use our models for Breast segmention of MRI Scans
2. Ease of Use: Anyone should be able to pull the container (without access issues), mount the input scans and get segmentations
3. Flexibility: Target users should be able to get outputs in different formats like dicom, nifti, nrrd etc
4. Scalability: Support for single point inference as well as batched inference

Approach and Plan

We already have trained model weights and a python module that provides an interface for segmentation using the aforesaid model. Next steps for us would be to:

• Publish models weights in Zenodo
• Wrap around segmentation module using Mhub core APIs
• Build and test the container
• Create documentation

Progress and Next Steps

1. Trained the Breast and tumor segmentation model
2. Wrote a segmentation / inference module using trained model weights
3. Added support for nifti and nrrd outputs

Illustrations

Coming soon

Background and References

Coming soon