AI X-Ray for 'COVID-19 Diagnosis'... 86% Accuracy

An example of a probability distribution feature map for COVID-19 obtained through a non-invasive COVID-19 diagnostic algorithm. While suspected areas rarely appeared in other lesions, in the case of COVID-19, the suspected regions are clearly highlighted in high resolution.

[Asia Economy Reporter Junho Hwang] Domestic researchers have developed a technology that can diagnose COVID-19 by analyzing chest X-ray images using artificial intelligence. This technology enables rapid and continuous diagnosis with an accuracy of about 86%. It is expected to become a diagnostic method that improves the efficiency of COVID-19 screening and treatment.

The Korea Advanced Institute of Science and Technology (KAIST) announced on the 25th that a research team led by Professor Yejongcheol from the Department of Bio and Brain Engineering developed a technology that dramatically improves the accuracy of COVID-19 diagnosis based on simple chest radiography images, which was introduced in the international academic journal 'IEEE Transactions on Medical Imaging.'

COVID-19 Diagnosis Using Artificial Intelligence

On the 24th, candidates are undergoing temperature checks to enter the venue for the SK Group's first half-year new college graduate recruitment written exam, the Comprehensive Competency Test (SKCT), held at Seokyeong University in Seongbuk-gu, Seoul. According to government COVID-19 prevention guidelines, candidates must wear gloves and masks during the exam. The exam results will be announced after the 1st of next month. Photo by Hyunmin Kim kimhyun81@

The research team addressed the limited amount of COVID-19 radiographic image data by distinguishing the heterogeneity between images through a consistent preprocessing process, enabling the identification of COVID-19-specific characteristic images. Then, using a 'local patch-based method,' they obtained various patch images from a single image to increase image diversity. Next, through a 'probabilistic feature map visualization method,' they created feature maps that highlight important areas necessary for COVID-19 diagnosis in high resolution. They confirmed that these maps corresponded with diagnostic radiological features.

To explain further, due to the scarcity of image data of lungs infected with COVID-19, they collected images of normal lungs, lungs infected with bacterial pneumonia, lungs infected with viral pneumonia, and lungs infected with COVID-19. The AI was then trained to distinguish features unique to COVID-19. To minimize the influence of image areas outside the lungs, the images were divided at the pixel level to find usable biomarkers within the lung region. This process yielded statistically significant data that could meaningfully differentiate lungs infected with COVID-19. Additionally, by cutting and training on different random regions within the lungs, the AI's accuracy was improved. The researchers created a feature map unique to COVID-19 infected lungs and confirmed that this map matched the diagnostic features of COVID-19.

Accuracy of 86%

Professor Yejong Cheol (from the left), Doctoral Candidate Yujin Oh, Doctoral Candidate Sangjun Park

The algorithm proposed by the research team showed a sensitivity of 86%, which is a 17% improvement over the 69% sensitivity of expert radiologists diagnosing COVID-19. In particular, when the feature maps highlighted important areas for COVID-19 diagnosis, multiple suspected lesion sites were effectively displayed in patients suspected of having COVID-19.

The research team expects that if this technology is further advanced, it will enable rapid and continuous COVID-19 diagnosis at any hospital in Korea at a low cost. Professor Yejongcheol said, "If the AI algorithm developed this time is used for patient screening and treatment, it will allow for rapid and continuous diagnosis of COVID-19 infection status, and by excluding patients with low probability, limited medical resources can be efficiently allocated to higher priority cases."

Currently, COVID-19 diagnostic tests typically use equipment based on reverse transcription polymerase chain reaction (RT-PCR). The accuracy of RT-PCR tests is over 90%. However, it takes a considerable amount of time to get results, and the cost is high to perform the test on all patients.

Although computed tomography (CT) scans also show relatively high accuracy, they take more time than general X-ray simple imaging tests and are difficult to use for screening due to the risk of equipment contamination by the virus.

This content was produced with the assistance of AI translation services.