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

Abstract

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.

Keywords: automatic treatment planning; dose prediction; multicriteria rating.

FIGURE 1

The automated planning pipeline. Using multicriteria rating to select the best‐performing deep learning model for predicting dose distribution. Isodose contours extracted from the predicted dose distribution are then used as objectives for plan optimization.

FIGURE 2

The dice similarity of isodose contours between predicted dose distributions from three deep learning models (U‐Net, DoseNet, and Transformer) and the ground truth.

FIGURE 3

Clinical acceptability of mean dosimetric parameters for automatic planning in the test dataset. (per‐protocol, minor variation, and major variation are evaluated according to Table S2).

FIGURE 4

Dose‐volume metrics statistics for auto‐plans and clinical plans, along with a paired‐sample t‐test assessment.

FIGURE 5

Dose‐volume histogram comparisons among the clinical plan (solid), IsoPlan (fine dashed), and DVH‐IsoPlan (bold dashed) for a randomly selected case from the test dataset.

FIGURE 6

Dice similarity (a) and Hausdorff distance (b) of clinical plans and auto‐plans.

FIGURE 7

Dose distribution of clinical plan (a), IsoPlan (b), DVH‐IsoPlan (c), and absolute dose difference between clinical plan and IsoPlan (d), clinical plan and DVH‐IsoPlan (e).

FIGURE 8

Dose gradient volume histograms of the predicted dose distributions/Isodose planning dose and the real dose for all 20 test set patients. x‐axis: gradient magnitude; y‐axis: volume/number of voxels MAE loss prediction and real dose (a); dice loss prediction and real dose (b); DVH loss prediction and real dose (c), hybrid loss prediction and real dose (d); IsoPlan dose and real dose (e); DVH‐IsoPlan dose and real dose (f).