Program Information
Machine Learning for the Prediction of Physician Edits to Clinical Auto-Contours in the Head-And-Neck
R McCarroll1*, J Yang1 , C Cardenas1 , P Balter1 , H Burger2 , S Dalvie2 , K Kisling1 , M Mejia3 , K Naidoo4 , C Nelson1 , D Followill1 , C Peterson1 , K Vorster5 , J Wetter2 , L Zhang1 , B Beadle6 , L Court1 , (1) UT MD Anderson Cancer Center, Houston, TX, (2) UCT Private Academic Hospital, Cape Town, Western Cape, (3) University of Santo Tomas Hospital - Benavides Cancer Institute, Manila, Metro Manila, (4) Universiteit Stellenbosch University and Tygerberg Hospital, Cape Town, Western Cape, (5) University of the Free State, Bloemfontein, Free State of South Africa, (6) Stanford University, Palo Alto, California
Presentations
TU-FG-605-9 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 605
Purpose: To validate and clinically implement auto-segmentation of normal tissues in the head-and-neck, including the development of a reinforced supervised machine learning tool for predicting those auto-contours which will require clinical editing.
Methods: For 128 patients, an in-house multi-atlas auto-segmentation tool was used to retrospectively contour 8 normal structures in the head-and-neck. These contours were assessed by a radiation oncologist for use in inverse planning and quantitatively compared to the clinically used structure using Dice and mean surface distance. After validation, the auto-contouring tool was introduced into the clinic and has since been used for 100+ patients. Clinically, the auto-contours are presented to the physician for editing prior to use in treatment planning. To differentiate those contours which require editing, a bagged tree classification model using contour features (e.g. volume, HU, shape, agreement with independent contouring, etc.) was developed. Internal 3-fold cross-validation was used to evaluate model performance. For 20 prospective patients, a comparison of edit prediction with the truth was made. To improve and refine the model, new, prospective patient data is being continuously added to the model and prediction accuracy is being tracked over time.
Results: On average, 92% of the auto-contoured structures [range: 73%(cochlea)-97%(mandible)] were rated acceptable for treatment planning. Average Dice between auto-contours and physician-drawn was 0.78 [range: 0.5(cochlea)-0.98(brain)]. Average mean surface distance was 2.9mm [range: 1.1mm(brain)-9.3mm(lungs)]. When implemented clinically, 50% of the auto-contours were used without edits. Internal 3-fold validation of classification models yielded an average sensitivity of 0.7 and specificity of 0.54. Prospectively tested on 20 patients, classification accuracy was 0.63.
Conclusion: Auto-contouring has been successfully implemented into the clinic for 8 normal tissues in the head and neck. A reinforced supervised machine learning tool has been developed to flag contours that are likely to need clinical editing, which may provide additionally time savings.
Funding Support, Disclosures, and Conflict of Interest: This work is funded by the NIH (UH2CA202665). Additional support from Varian Medical Systems and Mobius Medical Systems.
Contact Email: