Deep learning based image restoration algorithm for low-dose CT and mammography

Abstract

Radiography is a medical imaging technique that is essential when diagnosing a patient's condition. Among various types of radiography, there are computed tomography (CT), which uses X-rays to obtain three-dimensional images, and mammography, which can obtain two-dimensional images. These measure the intensity of X-rays which are reduced after passing through the patient, as the absorption of X-rays varies depending on the components that make up the body's various organs. Unlike surgery, it has the advantage of offering non-invasive viewing of the inside of the body, but it also runs the risk of radiation exposure. Therefore, active studies are underway to increase the safety of radiography. Typically, there are low-dose X-ray radiography techniques that reduce the number of photons in the X-ray beam. In addition, there are techniques that reduce the number of projections in X-ray CT or reduce the area irradiated by X-rays. This dissertation discusses image-restoration algorithms for low-dose medical radiography. Low-dose radiography has a low probability to be measured, but this increases the likelihood that noise will be added, and such noise will have a negative effect on the diagnosis of the patient's condition or the disease. The conventional method for improving the image quality of degraded low-dose radiographic images is the model-based iterative reconstruction (MBIR) method, which uses optimization techniques to obtain a clear image after the measurement acquisition process. However, it is computationally expensive given its long reconstruction time, and its performance is somewhat limited. In this study, we propose a deep convolutional neural network (CNN) to improve the performance of image-denoising algorithms. Deep CNNs are commonly used in the area of image processing, and their performance capabilities are significantly improved compared to existing methods. In the medical imaging area, deep CNNs are applied to studies of lesion segmentation or detection, though they are rarely applied to image restoration. First, we proposed a deep CNN structure which is suitable for low-dose CT and optimal learning conditions. It is difficult to understand the role of the network in existing CNN-based image processing methods, but fortunately we can understand its role by considering a theory related to deep convolutional framelets. With this theory, it is possible to improve the performance of an image-denoising algorithm by combining a frame-based iterative algorithm of which convergence is proved with a deep CNN. These techniques are deep CNNs based on supervised learning applicable only when the training dataset is perfectly matched between the input images and the target images. However, in actual clinical situations, two radiographs are required to obtain matched data, though this is difficult due to safety issues. It is also difficult to achieve perfect matching due to the patient's motion. With regard to cardiac imaging, it is impossible to obtain matched data due to cardiac motion. Therefore, image-denoising algorithms using unmatched data are necessary to obtain a clear image with various types of low-dose radiography. Here, we propose an image-denoising algorithm that utilizes cyclic consistency based on unsupervised learning and confirm that it is robust to the noise level and to the patient's motion.