Deep learning-based synthetic CT generation for MR-only radiotherapy of prostate cancer patients with 0.35T MRI linear accelerator

Purpose: To develop a deep learning model to generate synthetic CT for MR-only radiotherapy of prostate cancer patients treated with 0.35 T MRI linear accelerator. Materials and Methods: A U-NET convolutional neural network was developed to translate 0.35 T TRUFI MRI into electron density map using a novel cost function equalizing the contribution of various tissue types including fat, muscle, bone, and background air in training. The impact of training time, dataset size, image standardization, and data augmentation approaches was also quantified. Mean absolute error (MAE) between synthetic and planning CTs was calculated to measure the goodness of the model. Results: With 20 patients in training, our U-NET model has the potential to generate synthetic CT with a MAE of about 29.68 +/- 4.41, 16.34 +/- 2.67, 23.36 +/- 2.85, and 105.90 +/- 22.80 HU over the entire body, fat, muscle, and bone tissues, respectively. As expected, we found that the number of patients used for training and MAE are nonlinearly correlated. Data augmentation and our proposed loss function were effective to improve MAE by similar to 9% and similar to 18% in bony voxels, respectively. Increasing the training time and image standardization did not improve the accuracy of the model. Conclusion: A U-NET model has been developed and tested numerically to generate synthetic CT from 0.35T TRUFI MRI for MR-only radiotherapy of prostate cancer patients. Dosimetric evaluation using a large and independent dataset warrants the validity of the proposed model and the actual number of patients needed for the safe usage of the model in routine clinical workflow.

Automated summary

The study aimed to develop a deep learning model for generating synthetic CT for prostate cancer patients, showing that the U-NET model can generate synthetic CT of different tissue types during training; data augmentation and a specific loss function can improve model accuracy, but training time and image standardization do not significantly affect accuracy.