Program Information
Autosegmentation of Prostate Anatomy for Radiation Treatment Planning Using Machine Learning Models of Radiomic Features
MW Macomber1*, M Phillips1 , I Tarapov2 , A Nori3 , D Carter3 , L Le Folgoc3 , A Criminisi3 , MJ Nyflot1 , (1) University of Washington, Seattle, WA, (2) Microsoft Research, Redmond, WA, (3) Microsoft Research, Cambridge, UK
Presentations
SU-F-FS4-8 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: Four Seasons 4
Purpose: Autosegmentation of pelvic anatomy for radiotherapy planning would be useful for standardizing and expediting clinic workflow. We evaluated contours generated from machine learning models of radiomic features versus clinical contours for radiotherapy planning of the prostate.
Methods: We evaluated 52 treatment planning CTs for patients planned for prostate radiotherapy. 150 radiomic features (integral, scale-space, and rotation-invariant features) were extracted from the images. Machine learning models were trained via random decision forests which introduces randomization in the training for each tree via feature and data bagging. The models were used to classify image voxels as belonging to prostate, bladder, rectum, and right and left femoral heads on 35 training datasets and 17 testing datasets. The Dice coefficient was used to evaluate the accuracy of the automated contours versus the clinical contours, where 1 equals perfect agreement and 0 equals no overlap.
Results: Automated segmentation showed good agreement with clinical contours on training data. Dice coefficients for the prostate, bladder, rectum, right femoral head, and left femoral head were 0.836, 0.957, 0.815, 0.912, and 0.919 (n=35). Agreement with clinical contours was somewhat reduced but remained acceptable in the testing dataset. Two testing datasets contained atypical images of the prostate and bladder (bladder not filled per institutional protocol, prostate hypertrophy into the bladder) and were excluded for comparison for those contours. The resulting Dice coefficients for the prostate and bladder were 0.728 and 0.951 (n=15) while coefficients for rectum, right femoral head, and left femoral head were 0.710, 0.823, and 0.809 (n=17).
Conclusion: Machine learning models of radiomic features can generate contours of pelvic anatomy with reasonable agreement with physician contours on relatively small training datasets. Larger training and testing datasets are likely to improve model performance and robustness to image artifacts, indicating this method could be useful for automated treatment planning.
Funding Support, Disclosures, and Conflict of Interest: No financial support was provided. The software described is in development for potential commercialization by Microsoft.
Contact Email: