Use of artificial intelligence for detection of gastric lesions by magnetically controlled capsule endoscopy

Background and Aims Magnetically controlled capsule endoscopy (MCE) has become an efficient diagnostic modality for gastric diseases. We developed a novel automatic gastric lesion detection system to assist in diagnosis and reduce inter-physician variations. This study aimed to evaluate the diagnostic capability of the computer-aided detection system for MCE images. Methods We developed a novel automatic gastric lesion detection system based on a convolutional neural network (CNN) and faster region-based convolutional neural network (RCNN). A total of 1,023,955 MCE images from 797 patients were used to train and test the system. These images were divided into 7 categories (erosion, polyp, ulcer, submucosal tumor, xanthoma, normal mucosa, and invalid images). The primary endpoint was the sensitivity of the system. Results The system detected gastric focal lesions with 96.2% sensitivity (95% confidence interval [CI], 95.7%-96.5%), 76.2% specificity (95% CI, 75.97%-76.3%), 16.0% positive predictive value (95% CI, 15.7%-16.3%), 99.7% negative predictive value (95% CI, 99.74%-99.79%), and 77.1% accuracy (95% CI, 76.9%-77.3%) (sensitivity was 99.3% for erosions; 96.5% for polyps; 89.3% for ulcers; 87.2% for submucosal tumors; 90.6% for xanthomas; 67.8% for normal; and 96.1% for invalid images). Analysis of the receiver operating characteristic curve showed that the area under the curve for all positive images was 0.84. Image processing time was 44 milliseconds per image for the system and 0.38 ± 0.29 seconds per image for clinicians (P < .001). The kappa value of 2 times repeated reads was 1. Conclusions The CNN faster-RCNN-based diagnostic program system showed good performance in diagnosing gastric focal lesions in MCE images. Magnetically controlled capsule endoscopy (MCE) has become an efficient diagnostic modality for gastric diseases. We developed a novel automatic gastric lesion detection system to assist in diagnosis and reduce inter-physician variations. This study aimed to evaluate the diagnostic capability of the computer-aided detection system for MCE images. We developed a novel automatic gastric lesion detection system based on a convolutional neural network (CNN) and faster region-based convolutional neural network (RCNN). A total of 1,023,955 MCE images from 797 patients were used to train and test the system. These images were divided into 7 categories (erosion, polyp, ulcer, submucosal tumor, xanthoma, normal mucosa, and invalid images). The primary endpoint was the sensitivity of the system. The system detected gastric focal lesions with 96.2% sensitivity (95% confidence interval [CI], 95.7%-96.5%), 76.2% specificity (95% CI, 75.97%-76.3%), 16.0% positive predictive value (95% CI, 15.7%-16.3%), 99.7% negative predictive value (95% CI, 99.74%-99.79%), and 77.1% accuracy (95% CI, 76.9%-77.3%) (sensitivity was 99.3% for erosions; 96.5% for polyps; 89.3% for ulcers; 87.2% for submucosal tumors; 90.6% for xanthomas; 67.8% for normal; and 96.1% for invalid images). Analysis of the receiver operating characteristic curve showed that the area under the curve for all positive images was 0.84. Image processing time was 44 milliseconds per image for the system and 0.38 ± 0.29 seconds per image for clinicians (P < .001). The kappa value of 2 times repeated reads was 1. The CNN faster-RCNN-based diagnostic program system showed good performance in diagnosing gastric focal lesions in MCE images.