A Breast Auto-Contouring Model for Contour Verification in a Breast Radiotherapy Clinical Trial
H Baroudi*, C Nguyen, S Maroongroge, B Smith, K Hoffman, J Niedzielski, S Shaitelman, A Melancon, S Shete, T Whitaker, J Duryea, S Hernandez, M El Basha, R Mumme, T Netherton, L Court, UT MD Anderson Cancer Center, Houston, TX
Presentations
TU-I345-IePD-F3-2 (Tuesday, 7/12/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 3
Purpose: To develop tools to automatically verify the consistency of breast target contours in a clinical trial.
Methods: Breast CTV and regional lymph nodes were manually contoured and reviewed on CT scans of 104 post lumpectomy patients by radiation oncology residents. A deep learning model (nnU-Net) was then trained for left and right treatment sites using an 80/20 training/test ratio. Quantitative evaluation and physician review were conducted on 20 and 24 patients respectively. The auto-contouring model was then applied to radiotherapy plans from 7 patients who were enrolled in a clinical trial investigating hypofractionated breast radiotherapy. The consistency of the trial plans was assessed using the dose to 95% of the automatically generated CTV and PTV volumes (3mm expansion).
Results: Mean Dice Similarity Coefficient comparing automatic and manual contours were 0.92±0.02, 0.81±0.07, 0.80±0.07, 0.64±0.15, 0.75±0.10 and 0.68±0.08 for Breast CTV, Axilla levels I, II, III, supraclavicular nodes and internal mammary nodes, respectively. Somewhat low DSC for the lymph node CTVs was attributed to the small axial area/volume of these structures. Mean surface distance for all contours was lower than 4.8 mm. Physician review scored 63%/29%, 83%/17%, 100%/0%, 100%/0%,100%/0% and 100%/0% as use-as-is/requiring minor edits for Breast CTV, Axilla levels I, II, III, supraclavicular nodes and internal mammary nodes, respectively. When assessed on the previously-approved clinical trial plans, D95 was 99.6%±0.6%, 101.3%±0.8%, 87.7%±10.7%, 97.7%±10.8% for Breast CTV, combination of axilla lymph nodes with extension for aggressive treatment, supraclavicular nodes and internal mammary nodes, respectively. D95 for the corresponding PTVs was 99.9%±0.5%, 100.9%±2.8%, 96.0%±4.5% and 99.3%±9.9%, respectively.
Conclusion: Physician review indicated that the automatically generated contours are reliable. Dosimetric evaluation of these structures, when applied to approved clinical trial plans, demonstrated consistent coverage, indicating the likelihood that these tools can be used to flag inconsistent plans automatically.
Funding Support, Disclosures, and Conflict of Interest: Varian Medical Systems
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email