A Fully Automated Approach for Labeling, Contouring, and Treatment Planning of Vertebral Bodies, Including Automatic Verification to Prevent Errors
T Netherton*, C Nguyen, C Cardenas, C Chung, A Klopp, L Colbert, D Rhee, C Peterson, R Howell, P Balter, L Court, MD Anderson Cancer Center, Houston, TX
Presentations
SU-B-TRACK 6-1 (Sunday, 7/25/2021) 11:30 AM - 12:30 PM [Eastern Time (GMT-4)]
Purpose: To develop a fully-automated framework that labels vertebral levels, contours vertebral bodies, detects labeling and contouring errors, and generates treatment plans within 10 minutes. Additionally, to ensure that diagnostic and simulation CT-based auto-planning is safe, efficient, and flags potential mistreatments.
Methods: For labeling vertebral levels(C1-L5), a methodology was developed using two CNNs trained with independent patient populations(n1=897, n2=290). For contouring, UNet++ was trained, validated, and tested using diagnostic and simulation CT scans(n=220). For detection of contouring and labeling errors, 9 overlap, distance, and intensity metric features were calculated using 752 auto-contours from 56 CBCT, PET-CT, diagnostic-CT, and simulation-CT scans. Using these features as inputs, hyper-parameter optimization refined 9 machine-learning algorithms to detect errors. For planning, treatment plans were automatically generated based on auto-contour projections. To validate the entire framework, 60 treatment plans were generated on a new cohort of diagnostic and simulation scans varying in scan length, imaging protocol, presence of surgical implants, and metastatic burden. Clinical acceptability of the plans was evaluated by three radiation oncologists using a 5-point scale.
Results: Automatic labeling was accurate for 94% of scans and demonstrated mean localization error <2.2mm. Mean Dice-Similarity-Coefficient/95th-percentile Hausdorff distance was 85.0%/1.8mm(cervical), 90.3%/1.8mm(thoracic), and 93.7%/1.6mm(lumbar). A random forest model detected contouring and labeling errors across various CT scan types with precision-recall-AUC=0.82. Within the validation cohort, the random forest model flagged all contouring and labeling errors (11/11) within prescribed treatment regions, including errors from patient plans (6/6) with atypical anatomy (e.g. T13, L6). For patients with typical anatomy, radiation oncologists scored 98% of simulation CT- and 92% of diagnostic CT-based plans as clinically acceptable or needing minor edits. End-to-end treatment planning time was 7.0±2.0 minutes.
Conclusion: We demonstrated that this novel, end-to-end, fully automated process to label, contour, and plan was efficient, effective, and safe across various CT scan types.
Funding Support, Disclosures, and Conflict of Interest: I am a member of the Radiation Treatment Planning Assistant team at UT MD Anderson Funding for this work was provided by the Larry Deaven PhD Fellowship in Biomedical Sciences
Handouts
Keywords
Treatment Planning, Segmentation, Quality Assurance
Taxonomy
Contact Email