Author: Yu Chang, Mei Chen 👨🔬
Affiliation: Cancer Center, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Department of Radiation Oncology, Ruijin Hospital, Shanghai Jiaotong University School of Medicine 🌍
Purpose: Spot weights optimization, as a critical step in the proton therapy, is often time-consuming and labor-intensive. Deep learning, with its powerful learning and computational efficiency, can effectively accelerate various proton therapy tasks such as organ contouring, and dose calculation. This study is aim to exploring the feasibility of using deep learning method to predict spot weights of intensity-modulated proton therapy robust plan to accelerate the plan optimization process.
Methods: As an initial step, Single Field Optimization (SFO) robust plans were created for 120 prostate cancer patients from two angles (90° and 270°). The reference dose is calculated with all spot weights set to 1. A total of 240 data samples were generated, including 200 training samples, 20 validation samples, and 20 test samples. A 3D Attention Unet framework was employed, using the planned dose, spot mask, and reference dose as inputs, with the corresponding actual spot weights as the labels. The model was iteratively optimized using a Masked Mean Squared Error (MSE) loss function. Its performance was quantitatively assessed using the gamma passing rate and Dose Volume Histograms (DVH) metrics with robust analysis.
Results: For the 20 test samples, the model's spot weights achieved an average Masked MSE of 7.58 10-5. The reconstructed dose achieved an average gamma passing rate of 98.39±1.42% under the (2%, 2 mm) criterion and 99.73±0.32% under the (3%, 3 mm) criterion. The absolute errors for D95 and Dmean in target were only 0.021±0.014Gy and 0.014±0.012Gy, respectively. The absolute errors of Dmax for the rectum and bladder were 0.023±0.015 Gy and 0.022±0.019 Gy, respectively, while the Dmean errors were only 0.006±0.005 Gy and 0.007±0.008 Gy.
Conclusion: This study introduces a promising deep learning method that effectively predicts spot weights, thereby accelerating and taking the place of SFO plan robust optimization process in proton therapy.