Author: Steve B. Jiang, Mu-Han Lin, Yu-Chen Lin, Austen Matthew Maniscalco, Dan Nguyen, David Sher, Xinran Zhong 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab & Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Department of Radiation Oncology, UT Southwestern Medical Center, UT Dallas 🌍
Purpose:
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 both, resulting in a time-intensive, iterative optimization process. Existing dose prediction frameworks address only single plans, neglecting sequential boost complexities. Therefore, we propose a multi-plan dose prediction framework that models both plan-specific and plan-sum doses to provide more accurate dose estimates.
Methods:
We developed a U-Net-based Hybrid-Convolutional-Neural-Network(CNN) that processes CT images, OARs, PTVs, and dosimetric objectives 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 at three decoder levels to improve deep feature representation. Training loss is a vector of Mean Squared Error (MSE) for voxel-wise accuracy and Multi-Scale Structural Similarity Index (MS-SSIM) for regional coherence, and we used Jacobian descent via the torchjd package to ensure updates were beneficial to both loss components. This study used a site-agnostic dataset of 64 patients treated with sequential boost plans (38/6/20 training/validation/testing split).
Results:
The plan-sum model was compared to a baseline single-plan model, which omits plan-sum input data and only predicts plan doses. Models were trained to convergence, and Mean Absolute Percent Error (MAPE) was used to evaluate test performance, excluding voxels with doses <10% of prescription to avoid skew. The plan-sum model achieved statistically significant difference in MAPE (32.48±36.35%) compared to the single-plan model (42.26±66.93%). Both models achieved comparable SSIM values (0.96±0.03 vs. 0.95±0.05).
Conclusion:
Our multi-plan dose prediction framework improves voxel-wise accuracy by incorporating plan-sum information while maintaining strong perceptual consistency. This approach can streamline treatment planning by providing clinicians with an accurate, comprehensive strategy for dose allocation in sequential boost radiotherapy.