An automated planning optimization framework for cervical intensity-modulated radiation therapy using voxel dose prediction and adaptive parameter tuning

Graphical abstract

Highlights

• •Automated framework reduced rectum/bladder doses by 5.5 ± 2.2%/12.6 ± 9.4% at 45 Gy.
• •Parameter tuning ensured clinical compliance and 2.5 ± 2.1% extra organ sparing.
• •Automatic system enabled 100% autonomous planning without manual intervention.
• •Voxel-dose approach achieved 7.1 ± 4.6% lower rectum doses vs. dose-volume method.
• •Framework tolerated ±50% dose prediction errors with 0% plan violations (n = 27).

Abstract

Knowledge-based planning typically lacks mechanisms to manage prediction uncertainties. We developed an automated framework integrating voxel-dose prediction with adaptive optimization to overcome this limitation. For 27 cervical intensity-modulated radiation therapy (IMRT) cases, voxel-dose predictions initialized an end-to-end optimizer. A fuzzy-inference system dynamically adjusted objectives to correct clinical guidance violations and enhance sparing of organs of interest (OOIs). Results demonstrated significantly improved OOI sparing (e.g., rectum volume receiving 45 Gy was lower by 5.5 ± 2.2%, p < 0.005) while maintaining target coverage comparable to clinical plans. These findings indicate that the framework successfully generated high-quality IMRT plans by addressing prediction uncertainties through adaptive optimization.

1. Introduction

Treatment planning for intensity-modulated radiation therapy (IMRT) [1], [2] remains a complex, time-consuming process that requires iterative adjustments of dose goals for individual patients based on clinical protocols. This often results in lengthy planning times and outcome variability [3], [4]. Knowledge-based planning (KBP) [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17] has emerged as a promising alternative, using machine learning to predict patient-specific dose goals from historical plans. Early KBP models predicted dose-volume histograms (DVHs) or dose-volume indices (DVIs) [5], [6], [7], [8] but offered limited spatial information [7], [8], whereas recent voxel-based models provide more detailed and accurate three-dimensional (3D) doses [9], [10], [11], [12].

Despite these advancements, the direct integration of dose predictions into planning optimization remains underexplored. Existing approaches, such as DVH-based dose mimicking [13], voxel dose grid reconstruction [14], and isodose lines-based dose fall-off control [15], typically treat predicted dose as ideal with fixed goals. This strategy is often unstable due to predicted uncertainties, and lacks adaptive mechanisms to address clinical constraint violations.

Protocol-driven iterative planning (PDIP) offers a robust framework for adaptive optimization by encoding clinical guidelines into computable rules. Its core strength is real-time constraint enforcement; when violations are detected, PDIP autonomously recalibrates objectives and the priority of constraints using methods such as gradient-based tuning [16], [17], [18] or fuzzy logic [19]. However, the performance of PDIP is highly sensitive to its initial parameters, and inappropriate initialization can degrade final plan quality, which is a vulnerability that KBP's patient-specific predictions can directly address.

To synergize these complementary approaches, we propose a novel end-to-end optimization framework. Our method initializes the optimizer with a patient-specific, voxel-based KBP prediction. An iterative optimization engine then monitors DVI compliance using a fuzzy inference system to dynamically adjust objectives and weights when violations are detected. A generalized equivalent uniform dose (gEUD) objective is incorporated to enhance OOI sparing. This creates a self-stabilizing cycle: KBP provides an anatomically intelligent starting point, while automated parameter tuning ensures robust, protocol-compliant outcomes. We evaluated this framework through a multi-faceted analysis of its performance, component contributions, and robustness.

2. Materials and methods

2.1. Patient data and 3D dose prediction

A total of 136 clinical cervical IMRT plans were analyzed, Each plan was prescribed 45.0 Gy in 25 fractions to the planning target volume (PTV) using 7 coplanar beams (0°, 50°, 100°, 150°, 210°, 260°, 310°). All clinical plans were generated in Eclipse™(Varian Medical Systems, Palo Alto, CA) following the INTERTECC protocol (International Evaluation of Radiotherapy Technology Effectiveness in Cervical Cancer) [20], with OOIs including rectum, bladder, and intestine. This study was approved by the Ningbo No.2 Hospital Ethics Committee (approval number: PJ-NBEY-KY-2026-013-01).

A previously developed geometry-dosimetry model predicted dose distributions for OOIs [21]. The model is a feedforward backpropagation neural network with an input layer (9 nodes), three hidden layers (9 nodes each), and an output layer (1 node). It was trained and validated using 4-fold cross-validation, with 80% of cases for training and 20% for testing. Model performance was evaluated on 27 independent test cases using relative root-mean-square error (RMSE) for each region of interest (ROI). The inputs/output features and RMSE result are detailed in Supplementary Table S1 and S2.

2.2. Automated optimization framework

The core of our framework is a three-stage hierarchical fluence map optimization (FMO) model implemented and solved in MATLAB using the limited-memory Broyden–Fletcher–Goldfarb–Shanno sequential quadratic programming (L-BFGS-SQP) algorithm. A dose deposition matrix, pre-calculated using the quadrant-infinite-beam algorithm in the Computational Environment for Radiotherapy Research (CERR) platform [22], enabled efficient dose computation from the fluence map via a linear transformation. This standalone research setup was chosen for its flexibility in implementing our novel control logic. The overall workflow is illustrated in Fig. 1.

Fig. 1.

Workflow of the three-stage automated optimization framework. (a): Stage 1 generates an initial plan based on the predicted 3D dose; (b): Stage 2 iteratively refines the plan to satisfy clinical DVIs; (c): Stage 3 explores additional sparing for OOIs using gEUD objectives.

Initially, an FMO-based prediction-guided optimization stage generates an initial plan that approximates the predicted dose distribution derived from prior modeling. The objective function minimizes the sum of squared differences between calculated and predicted doses at the voxel level, thereby steering the optimization toward patient-specific dosimetric patterns. The detailed formulation follows established models [23], [24] and is provided in Supplementary Table S3.

Subsequently, the solution undergoes iterative refinement to ensure compliance with clinical DVIs, as listed in Supplementary Table S4. After each optimization, the dose distribution is evaluated against these constraints. Any violations trigger a fuzzy inference system (Section 2.3) to adjust the objective function weights for the relevant voxels. Supplementary Fig. S1 visually represents voxels violating different constraint types. For maximum DVIs (e.g., D_max < D^ref or V_D^ref < V^ref), violated voxels were those with a predicted dose > D^ref or within the range (D^ref, D_V^ref). For minimum DVIs (e.g., D_min > D^ref or D_v^ref > D^ref), violated voxels were those with a dose < D^ref or within the range (D_V^ref, D^ref). To prioritize target coverage, PTV constraints are addressed first; only after all PTV requirements are satisfied does the process proceed to OOI constraints, thus ensuring that OOI sparing do not compromise the already-met PTV requirements. This iterative process of optimization, evaluation, and parameter adjustment continues until all constraints are satisfied or a maximum iteration count is reached.

In the final stage, further exploration of OOI sparing is conducted while preserving strict adherence to the hard DVI constraints established in the previous stage. This is achieved by introducing gEUD objectives [25], which provide a convex function to efficiently drive dose reduction across entire OOIs. Starting from the Stage 2 output, the fuzzy inference system cautiously adjusts gEUD objective weights. The process terminates if any PTV DVIs falls below its tolerance threshold or upon reaching the maximum number of iterations.

2.3. Automated parameter adjustment mechanism

Objective and weighting factor adjustments were automated using a fuzzy inference system [19]. The system quantifies the deviations (e.g., small, medium, large) through membership functions, where a single deviation can belong to multiple categories with varying membership degrees (0–1). The system uses both input and output membership functions (Supplementary Fig. S2). The input is the percentage deviation between the expected dose and actual dose, and the output is the corresponding percentage adjustment to the objective/weighting. In this system, input is classified using graded membership degrees by the input membership function and then mapped to the output membership function using predefined fuzzy rules (i.e., large deviations prompt large adjustments). Each rule is evaluated to activate a fuzzy output, and all fuzzy outputs are aggregated into a single adjustment value using a center-of-gravity calculation (see example in Supplementary Fig. S3).

2.4. Automated plan generation and evaluation

The overall performance of the final Automated Plan was evaluated by direct comparison with the original Clinical Plans across all 27 cervical cancer cases. Plan quality was assessed by comparing dose distribution, DVH curves and 10 key DVIs. A paired t-test was used to evaluate the statistical significance of differences in DVIs, with a Bonferroni-corrected threshold of p < 0.005 (α = 0.05/10) to account for multiple comparisons across the 10 DVIs.

To dissect the contribution of each component within the three-stage framework, an ablation study was conducted comparing multiple planning strategies. The following plan types were generated and analyzed: Prediction-Only Plan: the result of direct optimization using the voxel-dose prediction (Stage 1). Prediction-Only Plan (DVH-based): the result of a traditional optimization using objectives from the DVH of the predicted 3D dose. This isolates the value of spatial guidance. DVI-Satisfied Plan: The plan after adaptive fine-tuning to meet all DVI constraints (Stage 2). Final Automated Plan: The final output after gEUD-based OOI sparing (Stage 3). Paired t-tests were used to compare these components across the 10 DVIs in three predefined pairwise contrasts. A Bonferroni-corrected threshold of p < 0.0017 (α = 0.05/30) was adopted to account for 30 total tests (3 comparisons × 10 DVIs).

The robustness of the framework to potential bias from the initial dose prediction was assessed via a sensitivity analysis. Intentional biases were introduced to simulate overly-conservative (dose predictions reduced by 10%, 30%, 50%) or underestimated (dose predictions increased by 10%, 30%, 50%) scenarios. These biased predictions were used as input for the full three-stage framework. The resulting Biased Automated Plans were compared to the unbiased baseline using paired t-tests, with a Bonferroni-corrected significance threshold of p < 8.3 × 10^-4 (α = 0.05/60) to adjust for 60 hypothesis tests (6 bias conditions × 10 DVIs).

3. Results

The final Automated Plan demonstrates steeper PTV dose gradients and superior dose sparing compared to Clinical Plan (Fig. 2a–d, DVIs provided in Supplementary Table S5), achieving significant absolute percentage-point reductions (p < 0.005) in rectum V_{30 Gy} (3.9 ± 2.3%) and V_{45 Gy} (5.5 ± 2.2%); bladder V_{30 Gy} (9.6 ± 5.4%) and V_{45 Gy} (12.6 ± 9.4%); and intestine V_{40 Gy} (4.7 ± 2.4%).

Fig. 2.

Plan quality and robustness of the automated framework. (a-c): Dose distribution and DVH comparisons between Automated Plan and Clinical Plan. The contoured structures: PTV (red line), rectum (green line), bladder (blue line), intestine (gray line). (d-f): DVH comparisons: (d) Ablation and baseline comparison: Plan evolution stages (Prediction-Only, DVI-Satisfied, Final Automated) and the initial prediction methods (Voxel-based vs. DVH-based); (e, f) Robustness test: Outcomes using a 30% tightened or relaxed dose prediction. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

In ablation analysis, the Automated Plan reduced rectum V_{30 Gy} by 1.9 ± 1.7% and bladder V_{30 Gy} by 2.8 ± 2.4% compared to the DVI-Satisfied Plan. The DVI-Satisfied Plan reduced intestine V_{40 Gy} by 2.5 ± 2.1% compared to the Prediction-Only Plan. The voxel-based Prediction-Only Plan resulted in lower rectum V_{45 Gy} (7.1 ± 4.6%) and bladder V_{45 Gy} (2.8 ± 1.1%) than the DVH-based Prediction-Only Plan (p < 0.0017), with comparable PTV coverage.

Under systematically tightened prediction bias (Fig. 2e–f, Supplementary Table S6), Automated Plans maintained PTV coverage while showing increases in rectum V_{45 Gy} (up to +11.6 ± 8.8%) and bladder V_{45 Gy} (up to +11.9 ± 9.3%). Intestine V_{40 Gy} decreased (up to −6.1 ± 1.4%; p < 8.3 × 10^-4) under tightened bias. With relaxed bias, rectum V_{30 Gy} decreased (up to −3.8 ± 3.3%) and bladder V_{30 Gy} decreased (up to −3.1 ± 2.7%), whereas intestine V_{40 Gy} increased (up to +1.3 ± 0.6%).

Across all bias scenarios, the final Automated Plan demonstrated lower bladder V_{30 Gy} (up to −4.6 ± 3.0%) compared to the DVI-Satisfied Plan. The DVI-Satisfied Plan showed variable performance: under tightened bias, rectum and bladder doses increased while PTV conformity improved; under relaxed bias, all OOI DVIs improved relative to the Prediction-Only Plan.

4. Discussion

This study developed and validated a novel end-to-end optimization framework integrating voxel-based KBP with a protocol-driven adaptive optimizer. Our results demonstrated automated generation of high-quality, clinically viable cervical cancer IMRT plans that significantly improved OOI sparing while robustly maintaining target coverage.

Existing KBP integration methods treat predicted doses as fixed targets, making them inherently unstable. While prior studies combined predicted objectives with clinical constraints [21], [26], their additive models remain parameter-sensitive. Our findings revealed a fundamental DVH-based paradigm limitation: voxel-based starting plans consistently outperformed DVH-based approaches (Fig. 2d, Supplementary Table S5), demonstrating that preserving spatial information provided a decisive advantage over collapsing 3D dose into DVH.

Our framework advanced beyond this paradigm by establishing a dynamic, prioritized feedback mechanism. The adaptive optimizer intelligently recalibrated objectives, relaxing biased predictions to ensure primary satisfaction of hard clinical constraints. Additionally, by incorporating gEUD minimization, we extended the scope of optimization to better spare OOIs. This created a more robust and fully automated process specifically designed to leverage the high-quality, spatially intelligent starting point provided by voxel-based KBP.

The ablation study confirmed the power of the multi-stage refinement strategy. The DVI satisfaction stage ensured clinical acceptability, while gEUD minimization stage provided additional OOI sparing. This sequential improvement was most effective when building upon the superior foundation of the voxel prediction. Crucially, the framework demonstrated robust performance, consistently producing clinically acceptable plans even when initialized with severely biased predictions. This finding underscored a pivotal conclusion: the adaptive logic was the fundamental guarantor of plan reliability and clinical feasibility, ensuring a robust baseline of quality. However, an accurate voxel-based KBP prediction provided a critical advantage by offering an anatomically intelligent starting point, which led to a more efficient optimization process and a superior result with more favorable trade-offs among ROIs.

This study had several limitations. As a proof-of-concept developed on an open-source platform, it did not incorporate machine-specific delivery constraints. Furthermore, validation was conducted on a single-anatomy cohort from a single institution. However, the underlying logic is not anatomy-specific. Future work will therefore focus on implementing this framework within a clinical treatment planning system and extending its application to other tumor sites. Nonetheless, this work successfully demonstrated that integrating voxel-based KBP with adaptive optimization enabled fully automated, clinically acceptable IMRT planning with significant OOI sparing, representing a meaningful step toward autonomous radiotherapy planning.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: