Deep Learning Driven Field Dose Prediction for Head and Neck Cancer Treated with Spot Scanning Proton Therapy

PurposeAccurate dose prediction is essential for automating radiotherapy planning. In spot scanning proton therapy (SSPT), dose evaluation is required at both the plan and field level. Evaluating individual treatment fields is critical to ensuring optimal beam angles are chosen to ensure target coverage and maximum organ-at-risk (OAR) sparing. Currently, however, no knowledge-based tools exist for predicting field-level doses for head and neck cancer (HNC) treated with SSPT. In this work, we aim to develop the first deep learning-based dose prediction model capable of field-level dose prediction for HNC treated with SSPT. MethodsA cohort of 62 HNC patients treated with SSPT was compiled for model development and evaluation. Collected patient data included treatment planning CTs, OAR masks, signed distance maps (SDMs), generated beam masks, and dose distributions. An encoder-decoder architecture enhanced with a cross-attention transformer bottleneck was used as the field prediction model. Comparison and ablation studies evaluated the models performance and determined the benefits of individual model components. Evaluation imaging metrics included mean absolute error, structural similarity index measure, and peak signal-to-noise ratio. Clinical performance was evaluated using dose-volume histogram metrics. ResultsThe best performing model from the ablation study was the full model using OAR masks, SDMs, generated beam masks and four-field dose prediction. The model outperformed the Distance Guided Dose Prediction (DGDP) and DeepLabV3 comparison models. The DGDP and DeepLabV3 comparison models had a mean validation set MAE performance of 1.268 Gy and 1.325 Gy, respectively, compared to our models mean validation set MAE performance of 0.949 Gy. The models final mean test set performance was MAE 1.024 Gy, SSIM 0.913, and PSNR 28.495 dB. ConclusionsWe developed a cross-attention transformer-enhanced deep learning model that accurately predicts per-field dose for HNC treated with SSPT, demonstrating superior performance over state-of-the-art models limited to plan-level dose prediction.