Oral abstract

343 CAN ARTIFICIAL INTELLIGENCE-BASED DIAGNOSTIC SYSTEM PERFORM DIFFERENTIAL DIAGNOSIS OF GASTRIC CANCER AND GASTRIC ULCER?

Background

We firstly in the world reported the usefulness of artificial intelligence (AI) system for detecting gastric cancers [1]. The Convolutional Neural Network (CNN) used in the previous research, which was named “original CNN” in the current study, was effective to detect abnormal lesions as suspiciously cancerous lesions but the accuracy for diagnosis was unsatisfactory.

Aim

The present study aimed to develop AI-based diagnostic system and evaluate its utility that can perform differential diagnosis of gastric cancer (GC) and gastric ulcer (GU).

Methods

We constructed the new CNN, which was named "advanced CNN” in current study, with making the original CNN developed. 5,193 endoscopic images of GUs were retrospectively obtained from the patients who underwent esophagogastroduodenoscopy (EGD) at our institution from January 2008 to December 2017. They included conventional white light imaging (WLI), chromoendoscopy with indigo carmine spraying, and narrow band imaging (NBI) and excluded any images with magnification. 4,453 images of 5,193

Result

The diagnostic time for analyzing 1479 test-set images by the advanced CNN was 38.0s. Regarding the accuracy for differential diagnosis of the advanced CNN using test-set mixed with GC and GU images, detection rate for GC and GU were 91.3% (675/739) and 94.5% (700/740). Sensitivity for GC and GU among the detected lesions were 98.5% (665/675) and 91.4% (640/700). Positive predictive value (PPV) for GC and GU among the detected lesions were 91.9% (665/724) and 98.9% (640/647).

Conclusion

AI-based diagnostic system with the advanced CNN achieved high detection rate and performance of differential diagnosis of gastric cancer and gastric ulcer.

1. Hirasawa T, Fujisaki J, Tada T, et al. Application of artificial intelligence using a convolutional neural network for detecting gastric cancer in endoscopic images. Gastric Cancer. 2018 Jul;21(4):653-660.