Dose prediction using a deep neural network for accelerated planning of rectal cancer radiotherapy

Purpose: To apply a deep neural network to predict dose distributions of rectal cancer patients for accelerated volume modulated arc technique (VMAT) planning.

Materials and methods: Computed tomography scans and approved VMAT plans together with Dose_approved of 187 patients treated from February 2018 to April 2019 were randomly selected for this retrospective study. The deep neural network DeepLabv3+ was applied for dose distribution prediction. A prior dose information-aided planning scheme was introduced. Prediction precision was evaluated by mean square error (MSE), normalized dose difference (δD), and dose-volume histogram (DVH) indices using a paired t test. Information-aided and experienced replanning were performed by 1-year and 6-year experienced dosimetrists, respectively. Replanning time and DVH indices were evaluated by two-way variance analysis.

Results: The DeepLabv3+ prediction results (Dose_DeepLabv3+) were all clinically acceptable. Taking Dose_approved as the baseline, the MSE was 0.001 and mean δD was 0.40% with an inter-quartile range of 0.079%-0.30% for Dose_DeepLabv3+. No significant differences were found for the planning target volume quantitative parameters between Dose_approved and Dose_DeepLabv3+, except for the conformality index. For the two-way variance analysis, a significantly different replanning time was found between the information-aided and experienced replanning with maximum time-saving of 15.76 min. Information-aided replans had the advantage of lower maximum dose, higher minimum dose, and lower homogeneity index, and the disadvantage of lower conformality index and higher machine unites with significant differences.

Conclusion: DeepLabv3+ successfully predicted rectal cancer dose distribution, and the predicted prior information helped save planning times for multi-level experienced dosimetrists.