Application of artificial intelligence technology based on convolutional neural network in early gastric cancer recognition

Abstract

Objective To construct and verify a convolutional neural network model for early gastric cancer recognition in order to improve the detection rate of early gastric cancer. Methods We collected the gastroscopy stock photos and gastroscopy videos in the database of the Endoscopy Center of our hospital from January 2016 to August 2020. A total of 5 496 photos from 928 patients were subjected, including early gastric cancer, benign lesions and normal pictures. The photos were randomly divided into training set (662 patients, 4 167 photos), and validation set (259 patients, 1 329 photos). The model was identified with 4 endoscopists, and finally the relevant Results were counted. Then, 458 photos of early gastric cancer identified by the model were randomly selected for the analysis of the accuracy of model positioning, and the overlap between the area of the lesion framed by the model and the area marked by the endoscopy experts was analyzed. Thirty-four cases of gastric endoscopy videos were used for real-time identification of the model. After the video lesion photos identified by the model were reviewed by endoscopic experts, the sensitivity index for the model to identify early gastric cancer was obtained. Results The recognition sensitivity and positive predictive value (PPV) of the model for early gastric cancer were 90.33% and 95.41%, respectively. The sensitivity and PPV for identifying benign lesions such as chronic superficial gastritis, gastric polyps and gastric ulcers were both over 80%. The sheet recognition time was 0.040±0.005 s. In terms of the detection rate and diagnosis time of early cancer, the model was better than the endoscopist group; after Chi-square test, it was better than the endoscopist group in the identification of early cancer lesions. In terms of accurate positioning of the model, there were 380 photos with an overlap of more than 60% of the model recognition, accounting for 82.97% of the total. In terms of morphology, the model showed most accurate for positioning uplifted lesions, and the number of picture of the overlap of different specifications (coincidence ≥60 %, coincidence ≥70%) were larger than flat and depressed lesions. In the video verification experiment, the model correctly identified 17 of the 19 early gastric cancer lesions with a sensitivity of 89.5%. The detection rate of early cancer lesions was in good consistency with the early cancer detection rate confirmed by biopsy pathology. Conclusion Our constructed convolutional neural network model presents high sensitivity and positive predictive value in the identification of early gastric cancer and 3 benign lesions (chronic superficial gastritis, gastric polyps, and gastric ulcer), and can accurately identify the location and margin of early cancer lesions. At the same time, it can also dynamically identify early gastric cancer and benign lesions, which can assist endoscopists in the actual clinical examination to increase the detection rate of early gastric cancer and improve the level of diagnosis.

Keywords

• convolutional neural network
• artificial intelligence
• gastric cancer