Abstract PO4-07-01: Deep learning guided radiation therapy machine selection via a multi-headed dose prediction model

Abstract

Abstract Background: Accelerated Partial Breast Irradiation (APBI) is gaining popularity for the treatment of appropriately selected breast cancer patients in lieu of whole breast irradiation regimens due to its comparable clinical outcomes and improved patient convenience and quality-of-life. Currently, several machines for delivery of APBI are available such as CyberKnife, Elekta Unity, Varian Ethos and Varian TrueBeam. Each treatment machine yields a unique distribution of radiation dose, so a given modality may afford particular advantages or liabilities based on a patient’s anatomy. Existing methods for identification of an ideal modality that best suits a patient’s unique anatomical features are generally lacking, manual, and resource intensive. To overcome this limitation and personalize modality selection, we propose a deep learning (DL) based multi-modality dose prediction model. The model takes a patient’s computed tomography (CT) scan, segmented planning target volumes (PTVs) and segmented organs at risk (OARs) as input, and outputs predicted dose distributions for all trained modalities. Differences may then be evaluated between each radiation treatment modality’s corresponding dose distribution and dose volumetric histograms (DVH) to select the optimal modality for each patient. Methods: Our dataset contains 16 partial breast patients, each with 4 unique treatment plans generated for 4 distinct modalities – CyberKnife, Elekta Unity, Varian Ethos and Varian Truebeam. 10, 2, and 4 patients were dedicated to the train, validation and test datasets respectively. Our model’s architecture is a U-net with a shared encoder and 4 independent dose decoders, in which each decoder predicts dose for one modality. Each training patient’s CT scan, PTVs, OARs, and a PTV distance map are input to the model during training, and the shared encoder shares the input data’s latent representations with each of the 4 decoders. The 4 distinct dose distributions from each training patient are used as labels for the 4 corresponding decoders to supervise model training. The mean squared error, mean squared log error, and DVH losses are minimized during model training. The model trained until validation loss converged. The 4 dose distributions were predicted in ~ 1 minute per test patient, yielding 16 dose distributions in the test dataset. Model performance across the test dataset was evaluated by calculation of the mean absolute percent error (MAPE) of dose differences within the patient’s body between the model’s predicted doses and label doses, relative to patient prescription dose. Results: The performance of the multi-modal dose prediction model was evaluated in terms of dose differences throughout each patient’s body across the 16 test dose distributions as compared to the actual planned doses. The average MAPE of the multi-modal dose predictions was 1.27 ± 0.38% of a patient’s prescription dose as compared to the actual planned dose across the test dataset. For a prescription dose of 3000 cGy, that corresponds to an average MAPE of 38.1 ± 11.4 cGy. Conclusion: Our multi-modality dose prediction model’s novelty lies in its rapid provision and comparison of accurate and detailed dosimetric data for each treatment modality. This model can empower treating radiation oncologists to make informed decisions in radiation therapy machine selection on a patient-by-patient basis, potentially optimizing radiation treatment plans by reduction of delivered dose to organs at risk. Table. Citation Format: Austen Maniscalco, Ezek Mathew, David Parsons, Asal Rahimi, Mona Arbab, Prasanna Alluri, Xingzhe Li, Mu-Han Lin, Steve Jiang, Dan Nguyen. Deep learning guided radiation therapy machine selection via a multi-headed dose prediction model [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO4-07-01.