A team of researchers at Seoul National University Hospital (SNUH) has developed a new diagnostic aid that can detect a small number of rare types of polyps based on artificial intelligence (AI) technology.
The hospital announced on Thursday that a joint research team led by Professor Lee Dong-heon of the Department of Radiology and Professor Kim Hyung-sin of Seoul National University Graduate School of Data Science has developed “ColonOOD,” a computerized diagnostic aid for colonoscopy, based on data from about 3,400 colonoscopies registered in four medical institutions in Korea and public datasets.
According to the latest National Cancer Registry statistics, colorectal cancer is the leading cancer type in Korea, accounting for the second-highest cancer incidence rate and the third-highest mortality rate. However, if polyps are diagnosed quickly and accurately through colonoscopy, the mortality rate of colorectal cancer can be reduced by up to 53 percent.
Colorectal polyps are primarily categorized into adenomatous polyps (high-risk) and hyperplastic polyps (low-risk). Computer-aided detection (CAD) systems are currently being introduced to colonoscopy to facilitate the quick and accurate diagnosis of these types.
However, most existing systems can only distinguish between the two main types of polyps, which limits their ability to detect infrequent or novel polyps. Therefore, a new diagnostic aid was needed with out-of-distribution (OOD) capabilities to detect a small number of untrained polyp types, providing clinicians with reliable results.
It is against this backdrop that the team developed ColonOOD, a diagnostic aid that automatically classifies the location and type of polyps based on colonoscopy images. The system can detect the distribution of minority types of polyps by learning the distribution of significant polyps. It also supports accurate decision-making by endoscopists by presenting the confidence level (high and low) of the classification results, which was not provided in the existing model, when classifying polyps.
That allows the endoscopist to categorize high-risk polyps initially. For other types, an additional analytic model is triggered to distinguish between low-risk polyps and potentially dangerous “minority type polyps.”
The researchers validated this classification performance on data from four medical institutions (Seoul National University Hospital, Gangnam Center, Asan Medical Center, Severance Hospital, and Ewha Womans University Seoul Hospital) and two public datasets.
They found that ColonOOD classified all polyps with up to 79.7 percent accuracy, and correctly detected up to 75.5 percent of all minority polyps.
In other words, ColonOOD automatically detects the location of polyps and identifies high-risk polyps, while polyps with uncertain classification are further analyzed to determine whether they are minority-type polyps, thereby increasing the reliability of the diagnosis, the researchers explained.
“This is the first study to integrate a small number of polyp detection modules into an existing AI-based colonoscopy diagnostic aid system, and we expect that clinicians can significantly improve diagnostic accuracy based on their confidence level by utilizing the predictive results of ColonOOD proposed in this study,” Professor Lee said. “ColonOOD, which reflects the actual clinical environment, is expected to be highly utilized in the medical field, and we will expand it to prospective studies and multicenter studies to verify its feasibility.”