Room: Track 5
Purpose: To quantify dosimetric differences from clinical(manual) versus deep-learning-based(automatic) clinical target volumes(CTVs) generated from fully-automated radiotherapy planning.
Methods/Materials: A convolutional-neural-network was previously trained to auto-delineate high-risk(CTV-high), intermediate-risk(CTV-int), and low-risk(CTV-low) CTVs using 285 oropharyngeal cancer patients. An independent-cohort of 32 patients were visually-inspected and scored based on a three-point scale (no/minor/major edits) by a head-and-neck radiation oncologist. Patients with “no-edits” CTV scores for all three CTV risk-levels were then selected to evaluate dosimetric differences between manual and auto-delineated target plans. Both plans were automatically-generated using the Radiation Planning Assistant. PTVs were prescribed to receive 70Gy/63Gy/57Gy. Dose differences in normal tissues between clinical and auto-delineated target plans were calculated. PTV coverage on the manual targets was assessed using doses from the auto-delineated target plans and compared to recently-reported inter-observer studies. Lastly, correlations between differences in coverage and overlap/distance metrics (automatic-vs-manual) were evaluated.
Results: Fifty-three-percent (17/32) cases had all three CTVs rated as “no-edits-needed” and only 1 case’s CTVs were scored as requiring “major-edits”. Doses to normal tissues were similar between manual and auto-delineated CTV plans with no statistical differences observed. The median percent difference in V100[%] between the manual and auto-delineated target plans was 4.2%(1.2-17.3%), 7.5%(3.7-20.1%), and 5.1%(1.1-20.5%) for PTV-high, PTV-int, and PTV-low, respectively. Distributions of differences in V100[%]/D98[Gy] were more consistent than target delineation inter-observer studies[2,3]. True-positive-fraction (overlap-metric) was found to have a strong linear-relationship with dosimetric differences in target coverage (average R²=0.824 for several metrics).
Conclusion: Physician-review showed that auto-delineated CTVs were of high-quality. As expected, small dosimetric differences were noticed on target coverage between the manual and auto-delineated plans. Target volume dosimetric differences could be closely estimated by calculating the true-positive-fraction between the manual and auto-delineated target volumes. CTV auto-delineation could have a significant impact in reducing inter-observer delineation variability, which could lead to more consistent patient outcomes.
Funding Support, Disclosures, and Conflict of Interest: Our research group receives funding from the NCI and Varian Medical Systems
Target Localization, Segmentation, Radiation Therapy