AI identifies tumors in colorectal cancer patients with 93% accuracy
Researchers from South Korea have used artificial intelligence (AI) to successfully identify tumors in histology images obtained from colorectal cancer (CRC) patients, sharing their findings in the Journal of Digital Imaging.
“Many recent studies have suggested using machine learning techniques to classify, localize, and segment tumor areas in histology images,” wrote author Hongjun Yoon of the National Cancer Center in Goyang, South Korea, and colleagues. “Deep neural networks (DNNs) are used extensively to extract and learn features of subjects; DNNs specifically adapted to image data, called convolutional neural networks (CNNs), can effectively classify or locate tumors; however, comparatively few studies have examined CRC using these techniques.”
Histology images were gathered from Korea’s National Cancer Center. A total of 30 CRC patients participated in the study. Images did need to be cropped due to memory issues with the computer’s graphics processing unit (GPU), an issue the others said was a result of their CNN’s complexity.
As the authors tested the CNN, they noted that its performance improved over time. At its peak, their system had an accuracy of more than 93 percent and a specificity of more than 92 percent. The best sensitivity they recorded was more than 95 percent. The experiment took the CNN more than 10 days to complete.
“Despite the excellent results obtained in this study, there is room for improvement because we restricted some variables for ease of computation,” the authors wrote. “First, we divided the histology image into smaller 256 pixels by 256 pixels images during preprocessing because of the inability of the GPU’s maximum capacity to handle larger images. However, the sliding-window technique, which enabled large images to be processed by breaking them into smaller images and efficiently concatenating all the features, could be used instead.”
Michael has more than 18 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.