(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
To automatically generate clinically acceptable field apertures and radiotherapy plans for rectal cancer.
Methods:
Standard rectal cancer radiation treatments use a three-field first-course (PA, opposed laterals) and a two-field cone down (opposed laterals). Clinically, these field apertures are primarily determined by bony landmarks and GTV extent as seen in DRRs. To mirror this field generation process automatically, a multi-stage DeepLabV3+ network was trained, validated and tested on 367 patients (221/73/73, respectively). First, a model was trained for the PA field using DRRs and projected GTVs as inputs. Second, a model was trained for the lateral fields using the output from the PA model together with DRRs and projected GTVs. This multi-stage approach incorporates additional information provided by the PA view to the lateral model. Third, a model was trained for the cone down fields using DRRs and projected GTVs as inputs. A script was written to automatically setup beams, and multi-leaf collimator, and perform final dose calculation.
Results:
The multi-stage process to predict field aperture and generate plans successfully for all 73 test patients. Target field prediction achieved average Dice coefficient of 0.96±0.02, 0.94±0.02, and 0.91±0.05 for PA, lateral fields and boost fields, respectively. The average Hausdorff distance was 1.15±0.59, 1.80±0.79, and 1.89±0.97 cm for PA, lateral fields and boost fields, respectively. The lateral model predicted more consistent field edges (superior and inferior) as the corresponding PA field, when using the multi-stage process compared to the predictions that did not use the additional inputs from PA fields. The entire planning process (including beam generation and dose calculation) took less than 13 min on average.
Conclusion:
We developed a multi-stage process mimicking the clinical workflow for rectal radiotherapy treatment planning. This automated process can successfully generate field apertures with high fidelity and can improve treatment planning efficiency.
Funding Support, Disclosures, and Conflict of Interest: Our research group receives funding from the NCI and Varian Medical Systems.
Keywords
Radiation Therapy, Treatment Planning, Segmentation
Taxonomy
Contact Email
