NA-MIC Project Weeks

Back to Projects List

Evaluation of AI methods for MRI segmentation on IDC data

Key Investigators

• Cosmin Ciausu (Brigham and Women's Hospital, US)
• Deepa Krishnaswamy (Brigham and Women's Hospital, US)
• Megha Kalia (Brigham and Women's Hospital, US)
• Andrey Fedorov (Brigham and Women's Hospital, US)

Presenter location: In-person

Project Description

We previously studied the application of a contrast-agnostic approach for MRI/CT abdomen organs segmentation, based on the generation of synthetic data. This synthetic data was then further used as a training set for a fully-supervised U-Net network. Since this study was performed, other methods aiming to segment MR abdominal organs have been published. Our goal is to evaluate these new methods on IDC MR abdominal-focused data and see how it compares to our method.

Objective

1. Evaluate performance of MR abdominal organs methods on IDC data.
2. Get feedback on our own method.

Approach and Plan

1. Select a subset of IDC MR abdominal-focused IDC data.
2. Create evaluation notebooks for newly published methods on this subset.
3. Compare to our method.

Progress and Next Steps

1. GitHub repo for colab notebooks for evaluation of MR segmentation methods
2. Look into methods like STAPLE for consensus of segmentations - WIP
3. Perform a comparison of the methods to ground truth - WIP

Illustrations

Comparison of MR segmentation methods on a subject from AMOS dataset:

• Top left = ground truth expert segmentations
• Top right = our approach
• Bottom left = TotalSegmentator
• Bottom middle = MRSegmentator
• Bottom right = our approach

Comparison of MR segmentation methods on a subject IDC TCGA-LIHC subject:

• 3D = our approach
• Left = our approach
• Middle = TotalSegmentator
• Right = MRSegmentator

Comparison of MR segmentations on a subject from TotalSegmentator: (ground truth in bold)

• Top row = our approach
• Middle row = TotalSegmentator
• Bottom row = MRSegmentator

Dice distributions between AI segmentations and expert annotations on AMOS22 MR training split.

Background and References

• Our method
  • Towards Automatic Abdominal MRI Organ Segmentation: Leveraging Synthesized Data Generated From CT Labels
• New published methods
  • MRSegmentator: Robust Multi-Modality Segmentation of 40 Classes in MRI and CT Sequences
  • TotalSegmentator MRI: Sequence-Independent Segmentation of 59 Anatomical Structures in MR images
  • MRISegmentator-Abdomen: A Fully Automated Multi-Organ and Structure Segmentation Tool for T1-weighted Abdominal MRI
  • TotalVibeSegmentator: Full Torso Segmentation for the NAKO and UK Biobank in Volumetric Interpolated Breath-hold Examination Body Images