Fujitsu Develops AI-Based Technology to Retrieve Similar Disease Cases in CT Inspections

26 Jun

Fujitsu Laboratories Ltd. has announced development of a technology to retrieve similar disease cases from a computed tomography (CT) database of previously taken images. The technology, jointly developed with Fujitsu R&D Center Co., Ltd., works by retrieving similar cases of abnormal shadows expanding in a three-dimensional manner.

Technologies already exist to retrieve similar cases based on CT images for such diseases as early-stage lung cancer, in which abnormal shadows are concentrated in one place. For diffuse lung diseases like pneumonia, however, in which abnormal shadows are spread throughout the organ in all directions, it has been necessary for doctors to reconfirm three-dimensional similarities, increasing the time needed to reach a conclusion.

Now Fujitsu Laboratories has developed an AI-based technology that can accurately retrieve similar cases in which abnormal shadows have spread in three dimensions. The technology automatically separates the complex interior of the organ into areas through image analysis, and uses machine learning to recognize abnormal shadow candidates in each area. By dividing up the organ spatially into periphery, core, top, bottom, left and right, and focusing on the spread of the abnormal shadows in each area, it becomes possible to view things in the same way doctors do when determining similarities for diagnosis. In joint research with Professor Kazuo Awai of the Department of Diagnostic Radiology, Institute and Graduate School of Biomedical Sciences, Hiroshima University, this technology was tested using real-world data, and the result was an accuracy rate of 85% in the top five retrievals among correct answers predetermined by doctors. This technology is expected to lead to increased efficiency in diagnostic tasks for doctors, and could reduce the time required to identify the correct diagnosis for cases in which identification previously took a great deal of time.

Going forward, Fujitsu Laboratories will conduct numerous field trials using CT images for a variety of cases, while additionally aiming to contribute to the increased efficiency of medical care by deploying this technology with related solutions from Fujitsu Limited.

Details of this technology will be announced at the Pattern Recognition and Media Understanding (PRMU) conference to be held by the Institute of Electronics, Information and Communication Engineers at Tohoku University (Sendai, Miyagi prefecture) on June 22-23.

Development Background

The number of images produced in imaging inspections for detecting diseases using CTs is increasing with the growing sophistication of imaging equipment, making an ever-increasing workload for doctors. This is particularly true because a significant percentage of the number of chest CT inspections consist of those for a group of diseases called diffuse lung diseases, in which abnormal shadows spread across the whole of the lungs, including numerous diseases such as interstitial pneumonia and emphysema. The interpretation and diagnosis based on these CT images requires a great deal of knowledge and experience, as well as a significant amount of time, and has become an issue for doctors. This has created demand for a technology that retrieves similar cases from the past with diagnosis and treatment information that can serve as a reference in the doctor’s decision making, in order to improve the efficiency of interpretation and diagnosis.

Issues

There are existing technologies that allow a doctor to specify an area of focus in a certain slice image, and retrieve other patients with similar slice images. These technologies have been useful when the abnormal shadows are concentrated in one place, as with early-stage lung cancer. In the case of diffuse lung diseases, however, in which the abnormal shadows are spread in all directions across the organ as a whole, retrievals employing this method could find cases that while appearing similar in certain slice images, would not necessarily look the same in three dimensions. To rule out such cases, doctors had to re-check the results to ensure their similarity in three dimensions, taking up a great deal of time (Figure 1).