BACKGROUND: This study aimed to prevent missing gastric cancer and point out low-quality images by developing a double-check support system (DCSS) for esophagogastroduodenoscopy (EGD) still images using artificial intelligence. METHODS: We extracted 12,977 still EGD images from 855 cases with cancer [821 with early gastric carcinoma (EGC) and 34 malignant lymphoma (ML)] and developed a lesion detection system using 10,994 images. The remaining images were used as a test dataset. Additional validation was performed using a new dataset containing 50 EGC and 1,200 non-GC images by comparing the interpretation of ten endoscopists (five trainees and five experts). Furthermore, we developed another system to detect low-quality images, which are not suitable for diagnosis, using 2198 images. RESULTS: In the validation of 1983 images from the 124 cancer cases, the DCSS diagnosed cancer with a sensitivity of 89.2%, positive predictive value (PPV) of 93.3%, and an accuracy of 83.3%. EGC was detected in 93.2% and ML in 92.5% of cases. Comparing with the endoscopists, sensitivity was significantly higher in the DCSS, and the average diagnostic time was significantly shorter using the DCSS than that by the trainees. The sensitivity, specificity, PPV, and accuracy in detecting low-quality images were 65.8%, 93.1%, 79.6%, and 85.2% for "Blur" and 57.8%, 91.7%, 82.2%, and 78.1% for "Mucus adhesion," respectively. CONCLUSIONS: The DCSS showed excellent capability in detecting lesions and pointing out low-quality images.
BACKGROUND: This study aimed to prevent missing gastric cancer and point out low-quality images by developing a double-check support system (DCSS) for esophagogastroduodenoscopy (EGD) still images using artificial intelligence. METHODS: We extracted 12,977 still EGD images from 855 cases with cancer [821 with early gastric carcinoma (EGC) and 34 malignant lymphoma (ML)] and developed a lesion detection system using 10,994 images. The remaining images were used as a test dataset. Additional validation was performed using a new dataset containing 50 EGC and 1,200 non-GC images by comparing the interpretation of ten endoscopists (five trainees and five experts). Furthermore, we developed another system to detect low-quality images, which are not suitable for diagnosis, using 2198 images. RESULTS: In the validation of 1983 images from the 124 cancer cases, the DCSS diagnosed cancer with a sensitivity of 89.2%, positive predictive value (PPV) of 93.3%, and an accuracy of 83.3%. EGC was detected in 93.2% and ML in 92.5% of cases. Comparing with the endoscopists, sensitivity was significantly higher in the DCSS, and the average diagnostic time was significantly shorter using the DCSS than that by the trainees. The sensitivity, specificity, PPV, and accuracy in detecting low-quality images were 65.8%, 93.1%, 79.6%, and 85.2% for "Blur" and 57.8%, 91.7%, 82.2%, and 78.1% for "Mucus adhesion," respectively. CONCLUSIONS: The DCSS showed excellent capability in detecting lesions and pointing out low-quality images.