DR-only Carbon-ion radiotherapy treatment planning via deep learning

Abstract

Purpose: To evaluate the feasibility of patient-specific digital radiography (DR)-only treatment planning for carbon ion radiotherapy in anthropomorphic thorax-and-abdomen phantom and head-and-neck patients.

Methods: The study was conducted on the anthropomorphic phantom and head-and-neck patients. We collected computed tomography (CT) and DR images of the phantom and cone beam CT (CBCT) and DR images of the patients, respectively. Two different deep neural networks were established to correlate the relationships between DR and digitally reconstructed radiograph (DRR) images, as well as DRR and CT images. The similarity between CT and predicted CT images was evaluated by computing the mean absolute error (MAE), root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), respectively. Dose calculations on the predicted CT images were compared against the true CT-based dose distributions for carbon-ion radiotherapy treatment planning with intensity-modulated pencil-beam spot scanning. Relative dose differences in the target volumes and organ-at-risks were computed and three-dimensional gamma analyses (3 mm, 3%) were performed.

Results: The average MAE, RMSE, PSNR and SSIM of the framework were 0.007, 0.144, 37.496 and 0.973, respectively. The average relative dose differences between the predicted CT- and CT-based dose distributions at the same carbon-ion irradiation settings for the phantom and the patients were <2% and ≤4%, respectively. The average gamma pass-rates were >98% for the predicted CT-based versus CT-based carbon ion plans of the phantom and the patients.

Conclusion: We have demonstrated the feasibility of a patient-specific DR-only treatment planning workflow for heavy ion radiotherapy by using deep learning approach.

Keywords: Deep learning; Gamma pass-rate; Predicted CT; Synthetic DRR.