Four-dimensional cone-beam computed tomography (4D-CBCT) can visualize moving tumors, thus the 4D-CBCT-based adaptive radiation therapy (ART) may improve the quality of radiation therapy. The aim of this study is to improve the quality of 4D-CBCT images using cycle-generative adversarial network (Cycle-GAN) and evaluate these images by a quantitative index. In this study, unpaired thoracic 4D-CBCT images and four-dimensional multislice computed tomography (4D-MSCT) images in 20 patients were used for training, and synthesis of 4D-CBCT (sCT) images with improved quality was tested in another 10 patients. The mean error (ME) and mean absolute errors (MAE) were calculated to assess CT number deviation, and peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were used to evaluate image similarity. The sCT image generated by our Cycle-GAN model effectively reduced artifacts on 4D-CBCT image. The ME and MAE were 46.5 and 61.9 in lung regions, whereas soft tissue and bone regions insufficiently restored CT number. Results of the SSIM and PSNR were significantly improved in the sCT image. The proposed Cycle-GAN method generates sCT images with a quality close to 4D-MSCT image, particularly in the lung region; however, anatomical regions with soft tissue and bone still require further improvement.
Radiation dermatitis is classified into 5 grades from 1, representing mild disease to 5, death due to adverse events based on the Common Terminology Criteria for Adverse Events used in clinical practice. However, it is based on visual assessment; therefore, there are issues that depend on individual experience and knowledge. We have generated artificial case images and constructed a hybrid generation method for a radiation dermatitis grading support system using deep learning to deal with the limitation posed by the relatively small number of cases used for learning. In this paper, we created a model using the recently proposed EfficientNet model. Additionally, we described the method of final classification based on the Bayesian estimation of images, which are graded differently by evaluators. The learning model using EfficientNet-B0 to B7 with different image resolution and data extension method conditions showed an overall accuracy of 86.4%. Moreover, the final grade judgment was performed with multiple EfficientNet models to resolve the ambiguity of grade judgment. The effectiveness of the proposed method was confirmed via an evaluation experiment using the maximum a posteriori estimation method based on the Bayesian’ theorem (Bayesian estimation).
This overview presents techniques for improving quality of magnetic resonance (MR) images using deep learning. This article classifies acquisition processes of MR as pulse-sequence generation, image reconstruction and post-processing. Applications of deep learning to those three processes are explained individually.
X-ray dark-field imaging (XDFI), which we have developed using synchrotron radiation light sources, can image soft tissues in three dimensions with high contrast comparable to that of stained pathological images. In this lecture, we introduce the imaging principle of XDFI (Part 1), the CT reconstruction method based on XDFI (Part 2), and the application of XDFI to medical research (Part 3), with the aim of familiarizing many readers with XDFI in an easy-to- understand manner. In this first article, an overview of XDFI and the imaging principle of XDFI are introduced, and then the mechanism of obtaining projection and tomographic images is described.