Prediction of 3D Radiation Dose Distribution for Breast Cancer Using A 3D U-Net Architecture

Abstract

Radiotherapy planning for breast cancer typically required a time-consuming, planner-dependent iterative process, motivating deep-learning-based dose prediction as a decision-support tool for consistent planning quality. This study developed and evaluated a three-dimensional (3D) U-Net predicting voxel-wise dose from a local patch (computed tomography, target/organ-at-risk masks, signed distance field), augmented by a global-context branch encoding a whole-volume representation (CT, planning target volume [PTV] signed distance field, beam isocenter distance map, and tumor laterality) that modulated the U-Net bottleneck via Feature-wise Linear Modulation (FiLM). The model was trained and validated on clinical DICOM-RT data of Indonesian breast cancer patients and evaluated on an independent test cohort of 24 patients using dose-volume histogram (DVH)-based dosimetric indices, the Homogeneity Index (HI), the Conformity Index (CI), the paired t-test, as well as qualitative isodose and DVH overlay comparisons. The predicted dose distributions achieved a mean absolute error for the planning target volume ranging from 0.644 Gy (Dmean) to 1.238 Gy (D95), an HI error of 0.045, and a CI error of 0.028. No statistically significant difference from the clinical plan was found for any PTV metric or for the heart (p > 0.05), and qualitative isodose and DVH comparisons corroborated this close concordance. Statistically significant but clinically moderate deviations were found only for the right lung, spinal cord, and trachea. These results demonstrated that the proposed FiLM-conditioned 3D U-Net achieved dose-prediction accuracy comparable to, and in several respects favorable against, previously reported international breast-cancer dose-prediction models, supporting its potential as a reliable quality-assurance tool for radiotherapy planning in clinical practice.