Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Restoration of Body Outside Field-Of-View On CT Images Using Cycle-Consistent Generative Adversarial Networks

K Arimoto1*, M Nakamura1, H Hirashima1, M Nakao1, T Mizowaki1, (1) Kyoto University


PO-GePV-M-141 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To restore body outside field-of-view (FOV) on computed tomography (CT) images using cycle-consistent generative adversarial networks (CycleGAN).

Methods: A total of 375 planning CT (planCT) datasets from 177 pancreatic cancer patients were included. Of them, 155 patients were scanned with breath-hold CT and 22 patients with four-dimensional CT. CT datasets with a limited FOV of 26 cm diameter (cropCT) were generated from planCT. The datasets were randomly divided into 240, 60, and 75 datasets for training, validation, and testing, respectively. Training datasets were augmented thirty-fold via translation and rotation in axial slices. As a result, 7440 CT datasets were used for training. The model for restoration of body outside FOV was built using CycleGAN. After input cropCT into the model, CT datasets including an entire body (predCT) were predicted. Training was conducted with a maximum of 1100 epochs, and the root mean square error (RMSE) between predCT and planCT was calculated in the validation dataset to determine the appropriate number of training epochs. Thereafter, RMSE, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) for the test dataset were calculated to assess the accuracy of prediction with the determined number of training epochs exhibiting the minimum value of RMSE within the bounding box tangent to the body on the planCT. DeepLearning BOX II (CPU: Core i7-9800X 8 core 3.80 GHz, GPU: NVIDIA Quadro GV100, RAM: 128 GB) was used for the calculations.

Results: The mean calculation time to predict the one CT dataset was about 30 s. At the 950 epochs, the RMSE was minimal in the validation datasets. The median RMSE, PSNR and SSIM for the test datasets were 21.6 (range, 12.0-35.0) HU, 33.3 (range 21.5-47.2) dB and 0.93 (range 0.82-0.99), respectively.

Conclusion: The proposed method has a capacity to restore the body outside FOV on CT images.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email