TSYS School Computer Scientist Yi Zhou Works to Improve Gastroscopy Imaging for Medical Diagnoses

In gastroscopy image diagnosis, the physiological structure of the stomach and peristalsis causes interference and blurred images, which increases the difficulty of lesion identification. Anchor-based object detection technology is a method that presets anchors with specific scales and aspect ratios in the image to determine the position and category of the target object. It can quickly and accurately identify various objects in complex scenes and thus provide an important basis for subsequent analysis and decision-making. However, anchor-based object detection techniques in gastroscopy image diagnosis suffers from limited localization accuracy, high computational cost, and scale sensitivity. Advanced anchor-free object detectors such as fully convolutional one-stage object detection (FCOS) were proposed to solve these problems. Unfortunately, for gastroscope image analysis, FCOS-based detectors have three drawbacks: (1) the centrality feature may not work well if the target center does not align with the center of the bounding box, (2) multi-scale feature fusion ignores the balance of non-adjacent feature map fusion, and (3) feature information is lost before performing classification tasks. To address these issues, a new study by TSYS School computer scientist Yi Zhou puts forward a gastric precancerous lesions detector (GPDet) featuring integrated dual center-ness fusion, a criss-cross-balanced feature pyramid, and fusion of fully connected layers. According to Zhou and researchers from Jinan University and South China University of Technology, dual center-ness fusion combines the semantic and location information of targets, enriching the feature content of center-ness. The criss-cross balanced feature pyramid focuses on the fusion of non-adjacent features, enhancing feature augmentation. The fully connected layer fusion reduces feature information loss in classification tasks, improving GPDet’s feature sensitivity. To assess the effectiveness of GPDet, Zhou and his coauthors conducted extensive tests on a collected gastric precancerous lesions dataset (GPD) along with two public datasets: HyperKvasir and Kvasir-Instrument. The findings, which are explained in greater detail in their article appearing in the current issue of the Journal of Supercomputing, clearly demonstrate that GPDet outperforms the existing solutions in terms of performance across these datasets, with its mean average precision (mAP) reaching 55.3%, 67.6%, and 74%, respectively.