An isodose-constrained automatic treatment planning strategy using a multicriteria predicted dose rating

Background

Previous knowledge-based planning studies have demonstrated the feasibility of predicting three-dimensional photon dose distributions and subsequently generating treatment plans. The steepness of dose fall-off represents a critical metric for clinical plan evaluation; however, dose fall-off similarity is frequently overlooked in dose prediction tasks. Our study introduces a novel automatic treatment planning methodology that specifically focuses on dose fall-off reconstruction for nasopharyngeal carcinoma (NPC).

Purpose

Our study aims to establish an innovative methodology for automatic treatment plan generation that leverages dose fall-off information derived from deep learning-predicted dose distributions. Additionally, we propose and validate a comprehensive multicriteria rating strategy for dose prediction.

Methods

We incorporated 120 nasopharyngeal cancer cases in this study, distributing them into training (n = 90), validation (n = 10), and testing (n = 20) cohorts. Three distinct dose prediction models were trained: U-Net, DoseNet, and Transformer. To determine the optimal dose prediction model, we developed a comprehensive multicriteria rating strategy that integrates mean absolute error, dose-volume histogram analysis, and isodose dice similarity coefficients. Based on these predictions, we implemented two automatic planning approaches: (1) IsoPlans, which extracts isodose lines from the predicted dose distribution to generate radiotherapy contours as optimization objectives and (2) DVH-IsoPlans, which enhances the first strategy by incorporating additional dose-volume constraints to further optimize treatment planning parameters.

Results

The multicriteria scores for the three dose prediction models (U-Net, DoseNet, and Transformer) were 0.85, 0.84, and 0.82, respectively. The dose prediction model achieved a minimum mean absolute error of 2.71 Gy. In our clinical validation, 4 of the 20 generated IsoPlans failed to meet clinical requirements, whereas all 20 DVH-IsoPlans successfully satisfied clinical requirements. The mean plan optimization time for the 20 test cases was significantly reduced from 870 to 560 s for IsoPlans and to 470 s for DVH-IsoPlans, representing a substantial reduction of 37.5% and 50.5%, respectively.

Conclusions

In this study, a multicriteria rating strategy is proposed which combines pixel-wise numerical evaluation, clinical parameter evaluation and physical dose fall-off evaluation in order to rate the dose prediction models. Moreover, an automated planning workflow has been developed, enabling the rapid generation of treatment plans based on the isodose structures of the predicted dose. A self-consistent dose prediction to automatic planning scheme based on isodose lines is proposed, which significantly reduces the time required for plan optimization.