关键词: 智能辅助决策, 电子胃镜, 上消化道疾病, 病灶检测, 深度学习

Abstract: With the acceleration of urbanization, industrialization and aging, upper gastrointestinal cancer has become one of the malignant tumors threatening the whole world.Disease screening using electronic gastroscope can detect potentially invasive, early stage, and precancerous patients in advance. The decision-making method for electronic gastroscopic image analysis based on deep learning is of great significance to improving the diagnostic efficiency and reducing doctors' working burden. However,in the process of decision-making for upper gastrointestinal diseases diagnosis, the existing methods are less involved in lesion detection and fine-grained classification and suffer from high computation time. There is a significant gap with the need of computer-aided decision-making in clinical practice. In addition, the traditional data augmentation strategies used in existing methods further reduce the actual performance of the decision-making methods. Therefore, a smart auxiliary diagnosis framework for upper gastrointestinal diseases based on electronic gastroscopic video is proposed in this paper. Firstly, the Conditional GAN is used to realize the enhancement of the original gastroscopic image, and then the k-Lconv module is designed for reducing computational complexity while improving accuracy of algorithm. Based on the YOLO-tiny algorithm, the upper gastrointestinal lesion detection algorithm Lconv-YOLO is developed. Lastly, the real clinical data from a Grade-Ⅲ Class-A hospital in China is used for test. The experimental results show that compared with other competitive methods, Lconv-YOLO can effectively improve the average accuracy of the upper gastrointestinal disease classification and lesion localization. The method shortens the time for detecting a frame of gastroscopic video on average to 6.73 ms, and the sensitivity and specificity reach 79.39% and 87.94%, respectively. It can meet the real-time high precision computer-aided diagnostic decision-making of video frames during electronic gastroscopy. The application of Lconv-YOLO can assist doctors in real-time diagnosis during gastroscopy and reduce the rate of missed diagnosis and misdiagnosis.

Key words: smart decision-making, electronic gastroscope, upper gastrointestinal diseases, lesion detection, deep learning