Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Auto-Segmentation of Organs at Risk for CT Thoracic Cases: A Single Institution's Experience

D Darkow1*, B Lee1, J Seger-Paisley2, R Woods2, U Duong2, A Houser2, J Paisley2, 3, M Butler1, A Kruzer1, A Nelson1, (1) MIM Software Inc. (2) New Hanover Regional Medical Center (3) Coastal Carolina Radiation Oncology

Presentations

PO-GePV-M-164 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: The purpose of this experiment is to examine one institution’s experience with a 3D Convolutional Neural Network (CNN) based automated contouring method.

Methods: A CNN trained on 658 expert-segmented images from multiple institutions was used to provide baseline contours for 129 patients undergoing thoracic radiation therapy. Five structures were included: heart, left/right lungs, esophagus, and spinal cord. Six dosimetrists edited those baseline contours to match their institutional standards. The Dice score (DSC) and Hausdorff distance (HD, mean/maximum) were calculated for each edited contour compared to the CNN contour. For each patient, the dosimetrists provided a self-reported scale of the edits required for each structure, choosing from “Minimal Edits”, “Moderate Edits”, “Not Usable”, and “N/A”. The “N/A” option was chosen if the structure was not required for the treatment plan or an automatically generated segmentation was not considered for use.

Results: The following average statistics were calculated for each structure (DSC, mean HD, maximum HD): Heart (0.92, 2.17, 14.89), Left Lung (0.98, 0.58, 12.61), Right Lung (0.98, 0.56, 13.16), Esophagus (0.76, 10.11, 44.79), Spinal Cord (0.86, 5.50, 35.71). Of the 292 structures for which the amount of edits were reported, 83% required minimal edits, 11% required moderate edits, and 6% were redrawn.

Conclusion: The CNN contours required minimal edits to yield clinically acceptable results, as demonstrated by the DSC, HD, and self-reported editing amounts. The esophagus and spinal cord yielded lower DSC and higher HD, (mean and standard deviation) due to variability in the superior/inferior extent of the structure that was contoured when using the CNN versus manual editing. In the future, we plan to investigate solutions to stylistic differences to improve the CNN performance with structures with variable extents.

Funding Support, Disclosures, and Conflict of Interest: Aaron Nelson is a part-owner of MIM Software Inc. Dan Darkow, Brandon Lee, Michael Butler, and Alexandria Kruzer are employees of MIM Software Inc.

ePosters

    Keywords

    Segmentation, Neural Network, Organs At Risk

    Taxonomy

    IM/TH- image Segmentation: CT

    Contact Email

    Share: