S1940
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
6 Department of Medical Physics, Kyonggi University, Suwon, Korea, Republic of
MLC 31.0±3.3 Gy), with a non-significant trend toward higher values in MLC (p = 0.060).
Purpose/Objective: Conventional dose prediction models in radiotherapy perform segmentation and dose prediction separately, requiring multiple inference steps and often failing to preserve anatomical consistency between predicted structures and dose distributions. This study aimed to evaluate the feasibility of a multi-task learning (MTL) framework that simultaneously predicts anatomical structures and dose distributions from planning CT images, providing anatomically consistent and efficient dose prediction for breast radiotherapy. Material/Methods: DICOM CT, RT Structure, and RT Dose datasets from 104 retrospective left-sided breast cancer patients treated with a simultaneous integrated boost (SIB) technique were used as model inputs. The multi-task attention adversarial network (MtAA-NET) was trained to perform a primary task of dose prediction and an auxiliary task of auto-segmentation [1]. The network takes planning CT images as input and simultaneously outputs the predicted dose distribution and segmentation masks of the PTV, boost, heart, and left lung. A shared encoder extracted common features, followed by task-specific decoders with independent hyperparameters. Attention maps guided the dose decoder toward anatomically relevant features, improving prediction accuracy and interpretability. Seventy-nine patients were used for training and twenty for validation. An independent test set of five patients was used to evaluate model performance. Model performance was evaluated both qualitatively and quantitatively against sequential prediction models (U-Net and cascade 3D U-Nets). Qualitative evaluation assessed the spatial agreement between predicted and reference structures and dose distributions, whereas quantitative evaluation employed the Dice similarity coefficient (DSC) for segmentation and the mean absolute error (MAE) for dose prediction. Results: For qualitative analysis, we confirmed that MTL- predicted dose maps exhibited improved spatial alignment along organ boundaries. The mean DSC values were 0.78 for PTV, 0.57 for boost, 0.90 for heart, and 0.94 for left lung. For target dose prediction, the MTL model showed the lowest accuracy among the three models, with a target coverage index (TCI) MAE of 0.22 for PTV and 0.82 for boost. For OAR dose prediction, the model achieved the second-highest accuracy, yielding an MAE of 0.06 for heart mean dose and a 1.07 Gy deviation for left lung V20Gy. Conclusion: This study demonstrated the feasibility of applying the MTL framework for anatomically consistent and
Conclusion: MLC is preferentially used for larger or more complex brain metastases without compromising conformity or homogeneity. Fixed collimators result in significantly less irradiation of healthy brain tissue, supporting their role as the optimal choice for small and well- defined lesions where achieving a highly steep dose fall-off is critical. Collimator selection should therefore be volume-driven, not quality-driven: both strategies maintain high plan quality. Keywords: SRS, robotic, brain Digital Poster 3559 Feasibility of a Multi-Task Learning Framework for Anatomically Consistent and Efficient Dose Prediction in Breast Radiotherapy Eun Jeong Heo 1,2 , Jae Choon Lee 1 , Song Heui Cho 1,3 , Dongyun Lee 4 , Kyung Hwan Chang 5 , Jang Bo Shim 3,6 , Nam Kwon Lee 1 , Chul Yong Kim 1 , Suk Lee 1 1 Department of Radiation Oncology, College of Medicine, Korea University, Seoul, Korea, Republic of. 2 Department of Medical Physics, Graduate School of Korea University, Sejong, Korea, Republic of. 3 Department of Radiation Oncology, Guro Hospital, Korea University Medical Center, Seoul, Korea, Republic of. 4 Department of Sales and CS, OncoSoft, Seoul, Korea, Republic of. 5 Department of Radiologic Science, Far East University, Seoul, Korea, Republic of.
Made with FlippingBook - Share PDF online