NA-MIC Project Weeks

Back to Projects List

Visualization and review of segmentations

Key Investigators

• Deepa Krishnaswamy (Brigham and Women's Hospital, USA)
• Vamsi Krishna Thiriveedhi (Brigham and Women's Hospital, USA)
• Cosmin Ciausu (Brigham and Women's Hospital, USA)
• Ron Kikinis (Brigham and Women's Hospital, USA)
• Andrey Fedorov (Brigham and Women's Hospital, USA)

Presenter location: In-person

Project Description

We have multiple sets of segmentation results that we would like to review with radiologists and clinicians. The first set of results are from running TotalSegmentator on patients from the National Lung Screening Trial (NLST). The second set includes results from work that Deepa and Cosmin are performing with training a model for abdominal MR/CT segmentation using synthesized data.

We would like to review this work, and showcase some results during project week.

Objective

1. Review results from TotalSegmentator on NLST patients. Use the Netter Atlas to learn about anatomy, and correlate this with the data that TotalSegmentator used for training, to further understand our results.
2. Review results from the abdominal segmentations on both MR and CT patients.

Approach and Plan

1. We will run TotalSegmentator on a sample of 1000 patients from NLST, convert to DICOM representation, and create OHIF links.
2. We will create DICOM SEG representations of the abdominal segmentations on IDC data, and create OHIF links.

Progress and Next Steps

General points:

1. How do we evaluate segmentations without ground truth?
2. What is also the best way to interpret our NLST segmentation results, using information we know about the data the pretrained model used? I think we should take a closer look at the training data to understand our results. For example clinical information, disease differences, etc.
3. How do we do outlier detection on large, heterogenous datasets?
4. What other radiomics features can we use besides volume? We are extracting the shape features, first order features, and general features from pyradiomics.
5. How do we curate patients/segments to be used for further analysis? For instance, not including patients that have incomplete segmentations – without having ground truth.
6. How do we make sure that we are correctly identifying cases where the laterality is incorrect?
7. Can we use information about the topology/atlas-based info to determine if segmentations are correct? Like we know left lower lobe is more inferior to left upper lobe.

Specific points:

1. What analysis can we do for the lung regions? What features besides the volume can we interpret? Can we take advantage of the NLST clinical tables? (smoking vs non smoking, etc).
2. For the vertebrae, are there heuristics we can do?
3. Are there heuristics that we develop for NLST that will work for SynthSeg evaluation?

** Work accomplished this weeek **

We had multiple sessions and discussions with Ron, where we:

• learned about anatomy
• tried to figure out what we could focus on for the NLST analysis and interpretation, and
• brainstormed how to develop better pipelines to view our data and segmentations

What we decided:

1. Focus on the liver and see if we can make correlations between liver health and lung cancer. For instance, fatty liver and cirrhosis and correlation with featuers we extract such as volume.
2. Brainstorm and develop better ways to quickly visualize our segmentations – using mrb file creation, CaseIterator, etc.

Illustrations

Example of TotalSegmentator analysis on an NLST patient in OHIF:

Example of liver analysis of NLST patient in Slicer:

Background and References

TotalSegmentator SynthSeg