Traditional gastroscopy, while effective in observing the stomach's condition, struggles to accurately assess the risk of gastric cancer. The development of gastric cancer typically progresses through a series of precancerous stages, including Helicobacter pylori infection, chronic gastritis, atrophic gastritis, and ultimately intestinal metaplasia. Currently, doctors must rely on endoscopic images and pathology reports to determine if a patient has precancerous lesions, a process that is time-consuming, labor-intensive, and costly, with a particular scarcity of resources in remote areas.
In light of this, Professor Yi-Chia Lee’s team at National Taiwan University Hospital has developed an "end-to-end system for the automatic identification of precancerous gastric lesions using artificial intelligence." Utilizing deep learning technology to analyze gastroscopy images, this system quickly identifies precancerous conditions such as atrophic gastritis and intestinal metaplasia, significantly enhancing examination efficiency.
System operation: A three-stage process
Image Preprocessing Enhances Diagnostic Accuracy
Before deep learning analysis, the system preprocesses the original color gastroscopy images by converting them to grayscale and using threshold segmentation to isolate the most relevant areas. It removes irrelevant information like black borders, patient data, and timestamps, then adjusts the images to a uniform resolution for model training.
Big data training yields 90% accuracy
The model is trained using a decade of historical gastric images and corresponding pathology reports from NTU Hospital. Additionally, it incorporates images and pathology data from Matsu’s ten-year gastric cancer prevention program for validation and testing. On the Matsu test set, the model achieved an accuracy of 99.9% for distinguishing gastric from non-gastric images, 98.2% for differentiating various gastric regions, and 90.1% and 88.6% accuracy for identifying atrophic gastritis and intestinal metaplasia, respectively.
Optimizing medical resource allocation in remote areas
Professor Lee emphasized that this system not only improves diagnostic accuracy but also aims to optimize medical resource allocation in remote areas. The team has collaborated with industry partners to develop a mobile PACS (Picture Archiving and Communication System) platform, enabling frontline doctors to use iPads to transmit images to NTU Hospital’s AI computing center for analysis. Results are quickly displayed on the tablet, highlighting risk areas. This system assists doctors in analysis and treatment planning, ensuring that limited medical resources are efficiently allocated to high-risk individuals, making a significant contribution to cancer prevention in remote regions.
Resource (mandarin):
