Transfer Learning for Fluence Map Prediction in Adrenal Stereotactic Body Radiation Therapy
W Wang1,2*, Y Sheng1, M Palta1, B Czito1, C Willett1, F Yin1,2, Q Wu1,2, Q J Wu1,2, Y Ge3, (1) Duke University Medical Center, Durham, NC, (2) Medical Physics Graduate Program, Duke University, Durham, NC, (3) University of North Carolina at Charlotte, Charlotte, NC
Presentations
TH-D-TRACK 6-4 (Thursday, 7/29/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Purpose: To predict fluence maps and automatically generate clinically acceptable plans for adrenal stereotactic body radiation therapy (SBRT) and to establish a deep learning framework that requires less demanding data volume using transfer learning (TL).
Methods: Due to the small number of adrenal SBRT cases, we developed a TL framework that leverages the treatment planning knowledge from a pancreas SBRT model. This framework consists of two convolutional neural networks (CNN), which sequentially predict individual beam doses (BD-CNN) and fluence maps (FM-CNN) for 9-beam intensity modulated radiation therapy (IMRT) plans. One hundred pancreas and 45 adrenal plans were retrospectively included. The two types of plans had different dose prescriptions and beam settings. A base model (including two CNNs) was trained with all 100 pancreas cases. The adrenal plan cohort were first split into training/validation/test sets with the ratio of 20/10/15 without TL. The base BD-CNN was re-trained using 5, 10, 15, and 20 adrenal cases with TL to produce candidate adrenal BD-CNN TL models. The base FM-CNN was directly used for adrenal cases without TL. The resulting AI plans were evaluated by PTV and OAR dosimetric endpoints to produce a percentage score with the clinical plans as the benchmark.
Results: Transfer learning significantly reduced the necessary training data size and training time. The TL models trained with 5/10/15/20 cases achieved the validation plan quality scores of 81.5/89.9/88.2/86.7, respectively. In contrast, a model using only adrenal training cases (maximum of 20) without TL only scored 78.2. For the test set, the 5/10/15/20-case TL models achieved the scores of 75.3/79.4/82.0/82.3.
Conclusion: It is feasible to use transfer learning to train the proposed fluence prediction framework, therefore reducing the required data size, training time and effort. Further, this strategy is capable of accommodating different beam geometries and dose prescriptions.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by the National Institutes of Health [grant number R01CA201212] and a master research grant from Varian Medical Systems.
Handouts
Keywords
Treatment Planning, Radiation Therapy
Taxonomy
Contact Email