Fully Convolutional Neural Network Model to Predict 3D Dose for Breast Radiotherapy
Presentations
SU-J-BRC-5 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]
Ballroom C
Purpose: To predict 3D radiation dose distributions for patients treated for breast cancer using a fully convolutional neural network (CNN). This dose prediction model serves as a proof-of-concept for a new approach to automated treatment planning for breast cancer.
Methods: To develop the model, 290 patients who were treated for left-sided breast cancer with a tangential-field technique were used. The input layers included contours of the tumor bed, ipsilateral lung, heart, and CT scan. The specific architecture that was used was a 3D U-Net using a mean squared error loss function. The model was trained using 5-fold cross validation on 244 patients and was evaluated on the remaining 46 test patients. To validate “in-field” dose prediction accuracy we set an evaluation threshold of >20% prescription dose and compared the predicted dose to the clinical dose distribution for each patient using mean squared error (MSE), average error (AE), and gamma analysis (5%/5mm criteria).
Results: The average MSE was 8.4%±3.4% and 9.5%±4.5% for validation/test sets and the AE was 8.3%±9.0% and 9.2%±9.3% for validation/test. The average gamma pass rate for our data was 72.38%. This compares well to results reported for 3D dose prediction for breast using other prediction approaches (78.68%).
Conclusion: We successfully developed a 3D model to predict the radiation dose for tangential-field treatment of breast cancer. The largest discrepancies were noted at the superior and inferior borders of the dose distribution. Investigating whether this is caused by model inaccuracy or clinical variability is the subject of future work. Additionally, we will test improving the model using various loss functions, such as those that weigh high dose regions and gradients, in addition to data augmentation. Overall, the model performed well on slices within the field, proving the concept of using these predictions for automated planning for breast.
Funding Support, Disclosures, and Conflict of Interest: Kelly Kisling acknowledges Honoraria and speaker fees from Varian Medical Systems. Kevin Moore acknowledges consulting fees and honoraria from Varian Medical Systems.
Keywords
Radiation Therapy, Breast, Treatment Planning
Taxonomy
TH- External Beam- Photons: IMRT/VMAT dose optimization algorithms
Contact Email