Application of virtual noncontrast CT generation technology from intravenous enhanced CT based on deep learning in proton radiotherapy

Abstract

Objective: The aim of this study is to generate virtual noncontrast (VNC) computed tomography (CT) from intravenous enhanced CT by using Unet convolutional neural network (CNN). The differences among enhanced, VNC, and noncontrast CT in proton dose calculation were compared. Methods: A total of 30 groups of CT images of patients who received enhanced and noncontrast CT were selected. Enhanced and noncontrast CT were registered. Among these patients, 20 groups of the CT images were chosen as the training set. Enhanced CT images were used as the input, and the corresponding noncontrast CT images were used as output to train the Unet neural network. The remaining 10 groups of CT images were chosen as the test set. VNC images were generated by the trained Unet neural network. The same proton radiotherapy plan for esophagus cancer was designed based on three images. Proton dose distributions in enhanced, VNC, and noncontrast CT were calculated. The relative dose differences in enhanced CT with VNC and noncontrast CT were analyzed. Results: The mean absolute error (MAE) of the CT values between enhanced and noncontrast CT was 32.3 ± 2.6 HU. The MAE of the CT values between VNC and noncontrast CT was 6.7 ± 1.3 HU. The mean values of the enhanced CT in the great vessel, heart, lung, liver, and spinal cord were significantly higher than those of noncontrast CT, he differences were 97, 83, 42, 40, and 10 HU, respectively. The mean values of the VNC CT showed no significant difference with noncontrast CT. The differences among enhanced, VNC, and noncontrast CT in terms of the average relative proton dose for clinical target volume (CTV), heart, great vessels, and lung were also investigated. The average relative proton doses of the enhanced CT for these organs were significantly lower than those of noncontrast CT. The largest difference was observed in the great vessel, while the differences in other organs were relatively small. The γ-passing rates of the enhanced and VNC CT were calculated by 2% dose difference and 2 mm distance to agreement. Results showed that the mean γ-passing rate of VNC CT was significantly higher than enhanced CT (p < 0.05). Conclusions: The proton radiotherapy design based on enhanced CT increased the range error, thereby resulting in calculation errors of the proton dose. Therefore, a technology that can be used to generate VNC CT from enhanced CT based on Unet neural network was proposed. The proton dose calculated based on VNC CT images was essentially consistent with that based on noncontrast CT.

Keywords

• Deep learning
• Convolutional neural network
• Enhanced CT
• Virtual noncontrast CT
• Proton radiotherapy