Artificial Intelligence-Based System Spots Stomach Cancer
Within a minute, the neural network correctly detected 92 percent of lesions in thousands of endoscopic images.
Stomach cancer is a leading cause of cancer-related deaths worldwide. Currently, doctors believe the most effective way to reduce stomach cancer mortality is through endoscopic detection at an early stage. Yet the standard endoscopic test for diagnosing gastric cancer is associated with a high false negative rate.
To tackle this problem, researchers recently constructed an artificial intelligence-based diagnostic system using a convolutional neural network that simulates the human brain. The study, which was published Jan. 15 in Gastric Cancer, is the first to evaluate the ability of a convolutional neural network to detect gastric cancer in endoscopic images, according to the authors.
The convolutional neural network was first trained on more than 13,000 endoscopic images of gastric cancer. During the test phase, the system took only 47 seconds to analyze a separate set of 2,296 stomach images collected from 69 patients with gastric cancer. The convolutional neural network showed an overall sensitivity of 92.2 percent, correctly diagnosing 71 of the 77 gastric cancer lesions.
However, only 71 of the 232 lesions diagnosed as gastric cancer were actually cancerous, resulting in a positive predictive value of 30.6 percent. According to the authors, this figure is clinically acceptable because false negatives are more problematic than false positives in diagnosing gastric cancer.
Because the procedure can be performed completely online, it might one day be used as a telemedicine tool to address the problem of insufficient numbers of endoscopists in rural areas as well as in developing countries, the researchers write.