Research on Multi-Classification System for Automatic Recognition of Gastroscope Images
Based on Transfer Learning
WANG Yue1,2,WANG Weidong1,ZHAO Lei2,ZHENG Tianlei2
1. College of Computer Science and Engineering, Jiangsu University of Science and Technology, Zhenjiang Jiangsu 212000, China丨
2. Department of Medical Equipment Management, The Affiliated Hospital of Xuzhou Medical University, Xuzhou Jiangsu 221000, China
Abstract:Objective To improve the image recognition accuracy of early gastric cancer by transfer learning. Methods Five types
of gastroscope images were colleted according to the concept of gastric precancerous lesions (PLGC), including 783 images of
early and advanced gastric cancer, 1042 images of gastric ulcer, 1143 images of chronic gastritis, 1096 images of gastric polyps, and
1763 images of normal gastroscope. These were divided into training set, validation set, and test according to the ratio of 6:2:2. By
comparing the results of the zero-training model ResNet34 with the fine-tuning migration models ResNet34 and VGG16. Results The
ResNet34 model based on transfer learning had the highest accuracy, 95.64% accuracy of validation set and 90.75% accuracy of test
set. Conclusion ResNet34 model can accurately realize the recognition of common gastroscope images, and has better generalization
and feature extraction capabilities for medical images with small data sets than traditional deep learning methods.
王跃1,2,王卫东1,赵蕾2,郑天雷2. 基于迁移学习的胃镜图像自动识别多分类系统的研究[J]. 中国医疗设备, 2021, 36(3): 81-84.
WANG Yue1,2,WANG Weidong1,ZHAO Lei2,ZHENG Tianlei2. Research on Multi-Classification System for Automatic Recognition of Gastroscope Images
Based on Transfer Learning. China Medical Devices, 2021, 36(3): 81-84.