Comparison of DVH-based machine learning and 3D convolutional neural network approaches for automated VMAT planning in head and neck cancer.

This study compared a dose-volume histogram-based machine learning (ML) approach with a three-dimensional dose distribution-based convolutional neural network (CNN) approach for volumetric-modulated arc therapy planning in head and neck cancer (HNC). Sixty-five patients who underwent whole-neck radiotherapy were retrospectively analyzed; 55 cases were used for model training and 10 for independent testing. Treatment plans generated by the CNN-based framework and a commercial ML-based planning system (RapidPlan) were evaluated using dose-volume indices (DVIs) and blinded qualitative scoring. In the DVI analysis, the ML-based plans achieved significantly higher target coverage than both the CNN-based and clinical plans. In contrast, the CNN-based plans maintained a mean error of less than 2% relative to the clinical plans, indicating close agreement with the clinical standard. No statistically significant differences in organs-at-risk dose metrics were observed among the three approaches. In the blinded qualitative evaluation, mean scores were 4.7 ± 0.56, 4.0 ± 1.07, and 2.7 ± 0.90 for the clinical, CNN-based, and ML-based plans, respectively, with the ML-based plans receiving significantly lower scores. These findings indicate that differences in prediction methodology and optimization strategy influence final plan quality, particularly with respect to spatial dose characteristics. Three-dimensional dose distribution-based prediction may provide clinical advantages for automated radiotherapy planning in HNC.