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

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Sequential boost radiotherapy (RT) presents challenges in dose allocation across multiple plans while protecting organs at risk (OARs).
• Current dose prediction models are limited to single plans, complicating sequential boost RT planning.
• The iterative optimization process for OAR sparing in sequential RT is time-intensive.

Purpose of the Study:

• To propose a multi-plan dose prediction framework for sequential boost RT that models individual and cumulative plan doses.
• To integrate the full RT course context for more efficient optimization objective establishment.
• To develop a versatile dose prediction approach adaptable to various RT scenarios.

Main Methods:

• A U-Net-based Hybrid Convolutional Neural Network (CNN) was developed to predict dose distributions for each plan and the plan-sum.
• The model incorporates pooling layers, skip connections, and a transformer bottleneck for global context.
• Training utilized a multi-objective loss function (MSE and MS-SSIM) with Jacobian descent on a site-agnostic dataset of 64 patients.

Main Results:

• The multi-plan model demonstrated statistically significant improvements over a single-plan model in both plan dose and plan-sum dose distributions.
• Key metrics showed lower Mean Absolute Error (MAE/Rx) and higher Structural Similarity Index Measure (SSIM) for the multi-plan approach.
• Specifically, MAE/Rx for plan-sum dose was 1.146 ± 0.174% vs. 1.525 ± 0.188% (p < 0.001), and SSIM was 0.972 ± 0.006 vs. 0.960 ± 0.006 (p < 0.001).

Conclusions:

• The proposed multi-plan dose prediction framework enhances accuracy and consistency by considering cumulative dose requirements across the full RT course.
• This approach streamlines treatment planning for sequential boost RT, offering clinicians a more accurate dose allocation strategy.
• The framework provides a foundation for a universal RT dose prediction model applicable to diverse treatment sites and fractionation schemes.