MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Evaluation of a Deep Learning Based Thoracic CT Segmentation Algorithm (DLSeg)

Q Chen1*, X Feng2 , K Qing3 , T Hunter1 , (1) University of Kentucky, Lexington, KY, (2) Carina Medical LLC, Lexington, KY, (3) University of Virginia, Charlottesville, VA

Presentations

(Sunday, 7/29/2018) 4:00 PM - 5:00 PM

Room: Exhibit Hall | Forum 2

Purpose: AAPM 2017 thoracic CT segmentation challenge reports that better than human contouring performance with reduced variability was achieved by several deep-learning based entries. However, the test dataset used came from same institutions that provided the training dataset using the same 4D scanning protocols. In addition, the reported final score was an average score over all organs and all cases, which can mask individual cases/organs that had much worse results. In this study, we tested the 2nd place model (DLSeg) to see 1) whether similar performance can be achieved on data from a different institution using different CT scanning protocol (regular non-4D) and 2) how much time saving can be achieved if this tool is used clinically.

Methods: Twenty-five lung cancer patients previously treated in our institution were randomly selected for this study. Using the previously drawn contour as reference, dice index, mean surface distance and 95% Hausdorff distance were computed for auto-segmented lungs, esophagus, heart and spinal-cord. A physician was invited to review the auto-segmentation and make changes if necessary. The time to review and edit the contours was recorded.

Results: The segmentation of lungs and spinal-cord on CTs from our institution were consistent with 2017 challenge result. Both achieved better-than-human performance in all three metrics used. Esophagus segmentation improved slightly due to minor changes done after the challenge. Heart segmentation was much worse. Both Esophagus and Heart segmentation were worse than human contour variation. The review and editing time required for each case was 7.5±2.4 minutes.

Conclusion: While there are still rooms for improvement for the current DLSeg model, it already provides significant time saving over traditional hour-long manual contouring process in a non-controlled environment.

Funding Support, Disclosures, and Conflict of Interest: We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research.

Keywords

CT, Segmentation, Lung

Taxonomy

IM/TH- image segmentation: CT

Contact Email