NA-MIC Project Weeks

Back to Projects List

Open Model for Anatomy Segmentation in Computer Tomography

Key Investigators

• Murong Xu (University of Zurich, Switzerland)
• Tamaz Amiranashvili (University of Zurich, Switzerland)
• Bjoern Menze (University of Zurich, Switzerland)
• Andrey Fedorov (Brigham and Women's Hospital, USA)

Project Description

We have developed a state-of-the-art automated segmentation model capable of identifying 167 anatomical structures in volumetric CT scans. This model has been trained on a combined dataset of more than 22,000 diverse, partially-annotated CT scans, setting a new benchmark in medical imaging. Our goal is to integrate this model into a 3D Slicer extension, making it widely available to the community.

Objective

1. Improve general user experience of the Slicer extension and finalize the development.
2. Prepare for performing large-scale inference on the IDC database.

Approach and Plan

1. Enhance user experience of our current prototype of the Slicer extension a. explore options for faster CPU-only inference b. add DICOM support c. incorporate SNOMED naming conventions
2. Finalize extension development (test extension on various OSs, writing tests)
3. Benchmark inference performance and prepare for large-scale inference on the NLST/IDC databases

Progress and Next Steps

Current Achievements:

• finalized Slicer extension UI and added DICOM support
• added support for SNOMED-CT naming conventions
• evaluated hardware requirements for inference on laptops
  • limited memory/CPU only: Trained smaller models
• initial process for working with IDC data: image retrieve, DICOM nifti conversion, restore

Next Steps:

• In progress: test on different OSs

Illustrations

Background and References

• Earlier PW41 project on this topic: OMAS CT: Open Model for Anatomy Segmentation in Computer Tomography
• Earlier work on mapping OMAS labels to SNOMED-CT: spreadsheet