MRI Diagnostic Accuracy Increased by AI … Cost and Time Reduced
Magnetic Resonance Imaging (MRI) is a diagnostic device that plays an important role in clinical diagnosis with X-rays, CT, and ultrasonograph. Unlike invasive methods in which a device for examination is inserted into the body, MRI provides high-resolution images by a noninvasive method, and thus is extensively used for observation and diagnosis of tumors and lesions. MRI can be used to produce T1 and T2 weighted images, and T1 contrast-enhanced images, which are necessary for diagnosis of brain tumors. The problem is that it is difficult to obtain all weighted images using an MRI performed in a hospital due to noise or artifacts. Different weighted images must be obtained according to the suspected disease to show, for example, specific cancer cells or blood. Therefore, patients may have to undergo imaging again, depending on their situations. For example, when a contrast medium is injected into the body to diagnose liver cancer, slight movement of the organs may hinder the acquisition of MRI images. In other cases, carious lesions, such as cerebral infarctions and cerebral hemorrhages, may be missed. To overcome these problems, Professor Jong Chul Ye’s group developed a technology of using AI to recover MRI images that had disappeared. The newly developed technology is drawing attention because it allows for accurate diagnosis and can reduce the cost and time needed to perform imaging repeatedly.
Systemized into ‘Collaborative Generative Adversarial Network’ technology
Recently, many studies have been conducted on technology to synthesize images by using the deep learning method called ‘Generative Adversarial Network (GAN).’ However, the application of the technology to the synthesis of MRI weighted images requires too many networks to be prepared and learned in advance. Professor Ye’s group independently developed the Collaborative GAN (CollaGAN) technology, in which the common characteristics spaces of several MRI weighted images are learned to solve the expandability problem. The technology was also used to develop a systematic method for answering the question of whether a certain contrast image may be generated or not. When images from several weighted images disappear in an arbitrary sequence and in an arbitrary number, the disappeared images are recovered by referring to the remaining images. After this recovery technology is learned, the clinical accuracy of the synthesized image is evaluated to automatically assess the relative importance of the weighted images. The experimental findings provide important guidance for the acquisitionprotocol design of MR in clinical environments. The results of the present study are significantly different from those of previous studies because computer vision, deep learning and MRI equipment technologies are combined to satisfy the demands of the medical workers.
“Patient inconvenience will decrease, and healthcare expense will be reduced”
Professor Ye commented, “Our technology shows excellent performance in generating missing images in the fields of not only healthcare but also computer vision, and thus will be applied to various areas.” Currently, transfer of the technology is being discussed for the application to clinical environment. With the COVID-19 pandemic, Professor Ye is looking for how to apply the technology to AI-based studies in precise diagnosis.
Furthermore, the images synthesized by AI can minimize patient inconvenience caused by repeatedly performed imaging, and reduce the overall healthcare cost by increasing the diagnostic accuracy. The results of the study were published in the January 2020 issues of Nature Machine Intelligence, and the article was included in the CVPR (2019) Best Paper Finalist.
Prof. Jong Chul Ye
2020 KI Annual Report