ABSTRACT
Objective:To investigate the pseudo-CT generation from cone beam CT (CBCT) by a deep learning method for the clinical need of adaptive radiotherapy.Methods:CBCT data from 74 prostate cancer patients collected by Varian On-Board Imager and their simulated positioning CT images were used for this study. The deformable registration was implemented by MIM software. And the data were randomly divided into the training set ( n=59) and test set ( n=15). U-net, Pix2PixGAN and CycleGAN were employed to learn the mapping from CBCT to simulated positioning CT. The evaluation indexes included mean absolute error (MAE), structural similarity index (SSIM) and peak signal to noise ratio (PSNR), with the deformed CT chosen as the reference. In addition, the quality of image was analyzed separately, including soft tissue resolution, image noise and artifacts, etc. Results:The MAE of images generated by U-net, Pix2PixGAN and CycleGAN were (29.4±16.1) HU, (37.1±14.4) HU and (34.3±17.3) HU, respectively. In terms of image quality, the images generated by U-net and Pix2PixGAN had excessive blur, resulting in image distortion; while the images generated by CycleGAN retained the CBCT image structure and improved the image quality.Conclusion:CycleGAN is able to effectively improve the quality of CBCT images, and has potential to be used in adaptive radiotherapy.
ABSTRACT
Objective:To investigate a time series deep learning model for respiratory motion prediction.Methods:Eighty pieces of respiratory motion data from lung cancer patients were used in this study. They were divided into a training set and a test set at a ratio of 8∶2. The Informer deep learning network was employed to predict the respiratory motions with a latency of about 600 ms. The model performance was evaluated based on normalized root mean square errors (nRMSEs) and relative root mean square errors (rRMSEs).Results:The Informer model outperformed the conventional multilayer perceptron (MLP) and long short-term memory (LSTM) models. The Informer model yielded an average nRMSE and rRMSE of 0.270 and 0.365, respectively, at a prediction time of 423 ms, and 0.380 and 0.379, respectively, at a prediction time of 615 ms.Conclusions:The Informer model performs well in the case of a longer prediction time and has potential application value for improving the effects of the real-time tracking technology.
ABSTRACT
Objective:To investigate the method of simulating low-dose CT (LDCT) images using routine dose level scanning mode to generate LDCT images with correspondence to the routine dose CT (RDCT) images in the training sets for deep learning model, which would be used for LDCT noise reduction.Methods:The CT images reconstructed by different algorithms in Philips CT Big Core had different noise levels, where the noise was larger with iDose 4 algorithm and lower with IMR(knowledge-based iterative model reconstruction)algorithm. A new method of replacing LDCT image with noise equivalent reconstructed image was proposed. The uniform module of CTP712 was scanned with the exposure of 250 mAs for RDCT, 35 mAs for LDCT. The images were reconstructed using IMR algorithm for LDCT images and iDose 4 algorithm at multiple noise reduction levels for RDCT images, respectively. The noise distribution of each image set was analyzed to find the noise equivalent images of LDCT. Then, RDCT images, those selected images were used for training cycle-consistent adversarial networks (CycleGAN)model, and the noise reduction ability of the proposed method on real LDCT images of phantom was tested. Results:The RDCT images generated with iDose 4 level 1 could substitute the LDCT images reconstructed with IMR algorithm. The radiation dose was reduced by 86% in low dose scanning. Using CycleGAN model, the noise reduction degree was 45% for uniform module, and 50%, 13%, 7% for CIRS-SBRT 038 phantom in the specific regions of brain, spinal cord, bone, respectively. Conclusions:Equivalent noise level reconstructed images could potentially serve as the alternative of LDCT images for deep learning network training to avoid additional radiation dose. The generated CT images had substantially reduced noise relative to that of LDCT.