3031 - Deep-Learning-Based Dose Prediction Compared with Multi-Criteria-Optimized Patient Doses and Benchmark Data for Prostate SBRT

Purpose/Objective(s): Deep-learning (DL)–based dose prediction models are increasingly implemented in radiotherapy treatment planning due to their ability to learn complex dose-anatomy relationships from large datasets, while multi-criteria-optimization (MCO) remains amongst the most effective methods for treatment planning. This study aims to demonstrate that DL-predicted dose distributions are non-inferior to human-generated, MCO treatment plans, and benchmark data.

Materials/Methods: Twenty prostate cancer patients were randomly selected from the LUND-PROBE open-access benchmark dataset for the comparison using 7 x 6.1Gy fractionation. For each patient, four dose distributions were evaluated: (1) DL-predicted, (2) VMAT optimized with MCO by an experienced medical physicist, (3) VMAT optimized with MCO by a medical physicist trainee, and (4) benchmark plan from the dataset. Dose-volume histograms (DVHs) were compared for planning target volume (PTV), rectum, bladder, femurs, penile bulb, and genitalia. Paired t-tests were used for statistical comparison.

Results: Comparison between (1) and (2) showed comparable PTV coverage with slightly smaller D_2% for (1). DVH metrics for the rectum and femurs were mainly similar between the methods. Penile bulb and genitalia presented significantly higher D_5% for (1) together with both bladder V_35Gy,20Gy metrics. However, all metrics were similar or better for (1) when compared to (4). Detailed dosimetric comparisons are presented in Table 1.

Conclusion: DL-based dose prediction provides valid target and OAR estimates proportional to human-generated plans in prostate radiotherapy. However, its performance depends on the training dataset, reflected in our study as higher dose predictions to the penile bulb and genitalia compared with an experienced MCO user. Hence, future validation and continuous model updates are necessary. In conclusion, our findings advocate the integration of DL-assisted planning into routine radiotherapy workflows but also highlight the need for transparent quality-assurance frameworks.