A generalizable 3D dose prediction across heterogeneous cohorts: A framework developed for the 2025 AAPM GDP‐HMM challenge

Abstract Background Manual radiotherapy treatment planning is labor‐intensive, subjective, and highly variable across planners and institutions. Deep learning has emerged as a promising approach for estimating achievable three‐dimensional (3D) dose distributions for a given plan (hereafter: dose prediction) by directly mapping patient anatomy and beam geometry to voxel‐wise dose. However, achieving robust generalization across diverse anatomical sites, prescriptions, and institutions remains a major challenge, particularly under computational constraints required for clinical deployment. Purpose We sought to develop and evaluate a generalizable, computationally efficient 3D dose prediction framework capable of spanning heterogeneous multi‐site and multi‐cohort datasets, as defined in the 2025 AAPM Generalizable Dose Prediction for Heterogeneous Multi‐Cohort and Multi‐Site (GDP‐HMM) Grand Challenge. Our goal was to design a model that not only achieved state‐of‐the‐art voxel‐wise accuracy but also maintained clinically meaningful dose‐volume and geometric characteristics across diverse cases. Methods We implemented a fully convolutional 3D dose‐prediction model adapted from the MedNeXt architecture, which was originally introduced as a high‐capacity convolutional backbone for medical image segmentation and is well suited to volumetric dense prediction tasks. The network accepted multimodal inputs comprising CT, PTV/OAR masks, beam and angle plates, and scalar plan metadata. A composite loss function combined voxel‐wise L 1 , body‐masked weighted L 1 , structure‐wise moment‐matching, and gradient‐based penalties to jointly preserve spatial fidelity and dosimetric accuracy. The model was trained on the fixed GDP‐HMM dataset consisting of 2876 training plans, 356 validation plans, and a held‐out test set of 498 plans using random 3D patches and evaluated by mean absolute error (MAE), homogeneity index (HI), and conformity indices (CI100%, CI50%). Ablation studies quantified the contribution of each loss component to overall performance. Statistical comparisons used paired Wilcoxon tests with false discovery rate correction. Results The proposed model achieved a mean voxel‐wise MAE of 2.04 Gy on the validation set, outperforming the challenge baseline (2.54 Gy). An ensemble with the baseline further improved MAE to 1.94 Gy on the validation set and 2.071 Gy on the held‐out test set, corresponding to the lowest test‐set MAE and first‐place overall ranking in the GDP‐HMM Challenge. Predicted dose distributions preserved key plan characteristics, with ΔHI, ΔCI 100% , and ΔCI 50% narrowly centered around zero, indicating consistent target coverage and dose conformity. Protocol‐specific evaluations for head‐and‐neck and lung cohorts demonstrated close alignment with reference plans, with no systematic degradation in OAR endpoints. Ablation experiments showed that removing gradient‐based terms increased MAE by 0.50 Gy, while excluding the moment‐matching term increased MAE by 0.06 Gy confirming that both gradient‐ and moment‐based constraints enhanced spatial and dosimetric fidelity. Conclusion We present a generalizable 3D dose prediction framework that achieves state‐of‐the‐art accuracy while satisfying stringent runtime requirements (< 3s per case on < 24 GB GPU memory). By integrating gradient‐aware and moment‐based constraints within a MedNeXt backbone, the model preserved clinical DVH endpoints and geometric plan features. These findings demonstrate that carefully engineered loss formulations are pivotal for cross‐institutional generalization. The framework offers a foundation for rapid dose prediction that could accelerate treatment planning, reduce inter‐planner variability, and enable intelligent automation in radiotherapy workflows.