Toward universal dose prediction: A multi-scale, multi-objective framework for sequential boost radiotherapy

Sequential boost radiotherapy (RT) poses a challenge in allocating dose across multiple plans while protecting organs at risk (OARs). Clinicians must decide whether OAR sparing should occur primarily in the initial plan, the boost plan(s), or all plans, resulting in a time-intensive, iterative optimization process.

Current dose prediction frameworks are limited to single plans and do not account for complexities introduced by sequential boosts. We propose a multi-plan dose prediction framework that models both individual plan doses and the cumulative plan-sum dose. By integrating the full course's context, this approach may help planners establish optimization objectives with fewer iterative adjustments. Ultimately, this framework aims to enable a versatile dose prediction approach adaptable to any RT course, regardless of treatment site, fractionation scheme, or the number of sequential plans.

We developed a U-Net-based Hybrid Convolutional Neural Network (CNN) that processes CT images, OARs, PTVs, and dosimetric goals to predict dose distributions for each plan and the plan-sum. It incorporates five pooling layers, skip connections, and a transformer bottleneck to capture global context. Deep supervision is applied in the decoder to encourage robust feature representation at deeper network layers, acting as a form of regularization. The single-plan model used for comparison differs from the multi-plan model in several key ways: (1) it only processes a single treatment plan at a time, (2) plan-sum dosimetric goals are re-scaled to reflect the plan fractions, (3) plan-sum PTVs are omitted, (4) it only predicts plan doses, and (5) plan-sum dose distributions were generated by summing plan dose predictions. Models were trained using a multi-objective loss vector of Mean Squared Error (MSE) for voxel-wise accuracy and Multi-Scale Structural Similarity Index (MS-SSIM) for regional coherence. Loss was averaged across all output levels to encourage global coherence. To improve training stability and ensure that parameter updates benefit both loss objectives, we used a Jacobian descent strategy via the TorchJD package. We collected a site-agnostic dataset of 64 patients that underwent sequential boost RT (38/6/20 training/validation/testing split).

The multi-plan model and single-plan model were trained to convergence. Evaluation metrics included Mean Absolute Error as a percent of prescription dose (MAE/Rx) and Structural Similarity Index Measure (SSIM). The multi-plan model achieved statistically significant differences versus the single-plan model in plan dose distributions with MAE/Rx (1.244 ± 0.151% vs. 1.650 ± 0.172%, p < 0.001) and SSIM (0.964 ± 0.005 vs. 0.944 ± 0.009, p < 0.001), and plan-sum dose distributions with MAE/Rx (1.146 ± 0.174% vs. 1.525 ± 0.188%, p < 0.001) and SSIM (0.972 ± 0.006 vs. 0.960 ± 0.006, p < 0.001).

Our multi-plan dose prediction framework improves voxel-wise accuracy and perceptual consistency by incorporating plan-sum information. Unlike traditional single-plan prediction models, our approach processes data from multiple treatment plans simultaneously, allowing it to consider cumulative dose requirements across the full treatment course. This approach can streamline treatment planning by providing clinicians with an accurate, comprehensive strategy for dose allocation in sequential boost RT. This framework lays the foundation for a universal RT dose prediction model capable of handling any fractionation scheme, disease site, or number of plans, which we plan to demonstrate in future work.