In recent years, the need for diagnosis using medical images from CT and MRI has been increasing. However, in medical fields around the world including Japan, there is a shortage of radiologists who are image diagnostic specialists, despite the prevalence of the machines. Due to this, depending on the region, the difficulty in quickly providing advanced image diagnosis has become a social issue. Also, with recent sophistication in diagnostic imaging equipment, which has rapidly increased the number of images that can be taken in a single scan, the growing diagnostic workload of radiologists has become an issue.
On the other hand, image diagnosis is extremely effective for early detection of serious diseases such as cancer, heart disease and cerebrovascular disease. It is an indispensable diagnostic method in improved preventive medicine. Therefore, the meaning of streamlining the workflow from producing images with the machines to diagnosing the images and creating an AI solution which can support high quality image diagnosis by radiologists, making it penetrate the medical field, is extremely significant.
With performance improvement of diagnostic imaging equipment such as CT and MRI, radiologists must carefully diagnose thousands of images taken from a single scan. This increases the burden in terms of time spent to create a diagnosis report for each patient.
There is a demand for definite detection of abnormalities from thousands of images.
Proof of concept (POC) for detecting abnormalities in kidneys has been conducted, and highly accurate detection of abnormalities was confirmed. Through continuous improvements in the future, it is expected that the diagnostic workload of radiologists will be further reduced, and to have the effect of supporting high quality image diagnosis.
Because it can detect various abnormalities, not limited to cancer, it is highly effective in preventing the overlooking of lesions through human error.
Through analysis of a patient's medical images with AI technology, possible disease areas are shown on a screen of a PACS system which is used for diagnosis, supporting accurate diagnoses by doctors. By providing AI diagnostic support information through interfaces used in the existing diagnostic processes, smooth collaboration between doctors and AI can be expected.
Can detect not only specific diseases but also various abnormalities in organs.
No reliance on CT imaging conditions such as differences in CT machine manufacturers and presence of contrast agents.
Can be introduced without major changes in existing hospital systems and workflow.
From these points, in addition to quick and accurate image diagnosis for highly urgent conditions, it can be expected to reduce the burden of radiologists and support diagnosis, even in preventive medicine including health checks.
Example of AI image diagnostic support solution
Recently, at University of Miyazaki Hospital, there has been a sudden increase in the number of image diagnostic cases, regardless of department. Also, with progress in diagnostic imaging equipment, there has been a great increase in the number of images taken in a single scan compared with the past.
Dr. Minako Azuma at Department of Radiology comments, "Because the CT images are reconstructed as easy-to-see 3D images, the number of image data that we radiologists have to diagnose is huge. Due to this, I feel that the time burden on doctors who have to create image diagnosis reports for each patient has greatly increased from the past."
Due to such actual conditions of image diagnosis, it is suggested that there is a potential risk of overlooking lesions through human error. Also, if a lesion unrelated to the main purpose of test appears on the images, radiologists must also discover it quickly and identify its nature and type.
"I often wondered if we could use AI to create a system that can detect lesions in images from a wider viewpoint, and prevent the slightest possibility of overlooking in advance." In order to realize Dr. Azuma's vision, the hospital looked into the possibility of image diagnostic support using cutting edge digital technology.
Dr. Minako Azuma
University of Miyazaki Hospital
Prior to this, NTT DATA had already developed a product which was to become the prototype for AI image diagnostic solution. This was developed using a medical image database of U.S. patients to construct algorithms to detect organ abnormalities using AI image recognition technology. Also, PoC was conducted at medical institutions.
"That's when I had the opportunity to talk with people from NTT DATA. After many discussions, we felt sure that they could customize the AI image diagnostic solution to accommodate the needs of patients and the workflow of doctors at our hospital, and decided upon joint development. Many vendors and researchers are involved in AI image diagnosis, but NTT DATA's AI is characteristic in the way that it has been designed especially with the aim of optimizing doctors' diagnostic efficiency, and gave us hope that they could work toward practical application. And in March 2019, in order to confirm if it could be applied to Japanese patients, following on from the U.S. and India, we started out with a PoC of AI image diagnosis for kidneys," recalls Dr. Azuma on the starting point of the project.
The AI used in the research at the University of Miyazaki had learned the characteristics of a diverse range of kidney disease patients using CT images of around 5,000 American patients, with cooperation from NTT DATA Services, which is NTT DATA's U.S. group company. In other words, when detecting abnormalities, diagnoses are made with algorithms based on U.S. patient data. So first, kidneys of Japanese patients at University of Miyazaki Hospital were examined and diagnosed through CT using this AI to verify if abnormalities can be automatically detected with the same accuracy as in the PoC of the U.S. Furthermore, at the hospital, CT examinations were conducted on patients being treated for various stages of kidney cancer in an attempt to verify its accuracy of identifying the condition as "cancer".
The development of the image diagnostic support solution using AI is driven with the cooperation of the following main players (still in progress).
|University of Miyazaki||Provides anonymized image data for verification. Offers medical knowledge in and participates in the joint research for interpreting AI results|
|NTTDATA||Develops/verifies algorithms, and supervises and promotes the overall project|
|NTT DATA Services||Provides AI learning data from its medical image archive in the U.S|
|MD.ai, Inc.||Provides technology that adds diagnostic information to medical images|
A healthcare AI solutions development specialist team from NTT DATA is participating in the development of algorithms. In defining the requirements for the image diagnostic support solution, algorithm design leader, Bun Dalia is not aiming to specialize in detecting a single disease, but has set the final goal as: "covering all diseases that can occur in the human body." When you take a look at the use of AI technology in the global image diagnostic field, there is a current trend that limits solutions to diagnosing specific organs and diseases. However, when doctors diagnose images, they do not assess a single disease, but are required to discover and diagnose various problems shown on the image. For that reason, although AI solutions for specific diseases are useful in preventing overlooking, she thought that there still remains an issue of not being able to contribute to diagnostic efficiency and time reduction. Also, she thought it was important that radiologists are able to apply AI image diagnosis without greatly changing their current workflow. NTT DATA's AI solution displays results on, not an original and new interface, but on a system called PACS, which is already being used in image diagnoses. Her target was a highly flexible solution which can be smoothly integrated with the hospital's existing systems and doctors' workflow.
In promoting the development project, the U.S. group company, NTT DATA Services, is in charge of providing learning data. It also improved the accuracy in detecting abnormalities by making AI learn through comprehensively selected CT images of a wide range of disease conditions from the company's cloud archive, Unified Clinical Archive (UCA), which stores over 19 billion medical images. Along with the progress in PoC, they continue to accumulate CT images of U.S., Indian and Japanese patients as AI learning data.
We are also collaborating with a U.S. startup company, MD.ai, Inc., which possesses original technology which adds diagnostic information to medical images. The radiologist at the company, Anouk Stein, MD, spoke about looking ahead to future issues. "There is a possibility that solutions using AI will be useful in treating patients in the future at various medical institutions around the world. In order to make further progress in medical AI, it is necessary that we repeatedly exchange information between industry and academia, and form a consensus."
NTT DATA Technology and Innovation General Headquarters
On receiving the results from the PoC which was conducted over 6 months at the University of Miyazaki Hospital, Dr. Azuma had a positive opinion and stated, "Even in image diagnoses of Japanese patients, AI detected various kidney abnormalities with higher than expected accuracy and I felt we can hold hope for the future." The solution has shown that it can detect various abnormalities, from diseases such as kidney cancer, kidney stones and hydronephrosis to cysts and benign tumors at the same level of accuracy as in the PoC conducted for U.S. patients. It can be said that, from the series of results, we have confirmed that the AI image diagnostic support solution can be applied to patients of different races and environments from multiple countries and regions.
Also, in regards to diagnostic accuracy for kidney cancer, an overall accuracy rate of 89% has been achieved. A high diagnostic accuracy rate has been confirmed with results of 82.00% for "sensitivity," which indicates that only few have been overlooked, 95.00% for "specificity," showing a low level of misjudging healthy patients as having cancer, and 94.60% for "precision rate," which indicates the rate of patients diagnosed with cancer to actually have cancer.
Dr. Azuma spoke of the outlook for the project in the following way. "AI image diagnosis at the moment is basically restricted to 'detecting abnormalities.' But as a target for the future, I want to be able to get to the stage where the nature and type of lesions, and degree of spreading can be distinguished in detail from the images. If we can identify the disease name with AI, it will be more beneficial to patients in terms of early detection, and will greatly relieve the burden on radiologists. I also think it will lead to a reduction in the risk of overlooking abnormalities."
For example, if we can predict the nature, such as a tumor being benign or malignant, at the stage of image diagnosis, radiologists can inform this to the attending physician as preoperative information, making it possible for the physician to make an appropriate explanation to the patient, decide on the surgery approach, and provide high quality post-surgery follow ups. "Eventually, I want to aim for AI image diagnosis which can be applied for whole body examinations. If we can realize this, I believe that doctors will be able to, in cooperation with AI, make speedy and accurate diagnosis, and we will be able to provide a beneficial medical service for patients in both acute-phase medical care and preventive medicine." (Dr. Azuma)
NTT DATA will continue to work closely with the University of Miyazaki Hospital, making improvements towards the commercialization of AI image diagnostic support solutions. We are already in the second stage of the PoC and started the development/verification of algorithms which identify disease names from detected abnormalities. Furthermore, by the end of FY2020, we plan to verify the effects, such as how much workload can actually be reduced for doctors at the hospital.
|Client name||University of Miyazaki (Faculty of Medicine Hospital)|
|Address||5200 Kihara, Kiyotake-cho, Miyazaki-shi, Miyazaki|
|Established||October 31, 1977 Established as the Hospital of the Faculty of Medicine, University of Miyazaki|
|Hospital outline||A core medical institution in Miyazaki Prefecture with 29 departments and a total of 632 beds.
Cooperates with regional medical institutions, practices advanced medicine and is involved in training excellent medical human resources.