An Early Gastric Cancer Detection Method Based on DETR Multi-scale Feature Fusion

At present, the incidence and mortality of gastric cancer rank in the top five of all cancers, which seriously endangers people’s life and health, and the detection of early gastric cancer can greatly improve the survival rate of gastric cancer patients. Based on the features of early gastric cancer in gastroscopy, this paper studies the intelligent detection of early gastric cancer by using object detection technology. Detection Transformer (DETR) is a mainstream object detection framework at present, it shows good detection effect on COCO data set, but it is rarely used in medical image object detection. Through in-depth study of DETR model, we find that DETR has shortcomings in multi-object and smaller object detection, which will affect the effect of detection. So, we propose a multi-scale feature fusion method for early gastric cancer detection (MF-DETR) based on the DETR model. The main improvements include multi-scale feature fusion module and hybrid encoder design. We rely on early gastric cancer cases provided by Chongqing University Cancer Hospital, and then enhance the data to form 2062 images as an early gastric cancer data set. Through experiments, MF-DETR can achieve 97.79% AP
50
, 92.31% precision, and 93.91% recall rate. Compared with DETR, it has 2.68% improvement in $\mathrm{AP}_{50}, 1.12 \%$ improvement in precision, and 3.91% improvement in recall. Finally, the effectiveness of multi-scale feature fusion design and hybrid encoder design are verified by ablation experiments.