딥러닝 기반 위내시경 데이터 해부학적 위치 분류 모델 개발 연구

Abstract

Gastric cancer is one of the main causes of cancer and death in Korea. Gastroscopy (EGD) is known as an effective gastric cancer screening technique. However, the quality of medical services is deteriorating as the doctor's workload increases due to non-detection of lesions caused by blind spots and dependence on treatment. Therefore, we conduct research on the development of a deep learning-based gastroscopy data anatomical position classification model for medical staff support. In this study, data were collected from about 604 patients who underwent gastroscopy through Gachon University Gil Hospital. Labeling of the data was classified based on anatomical location. This study is based on 6 categories consisting of a gastric cardia, gastric upper body, gastric middle body, gastric lower body, gastric angle, and gastric antrum and 4 categories that integrated the upper, middle and lower parts of the gastric body. We developed gastroscopy anatomical classification models using InceptionV3, and Inception-ResNetV2. In the case of 6 category classification, the InceptionV3 model was the highest F1-score with 62.07% F1-score. For 4 category classification, the Inception-ResNetV2 model was the highest F1-score of 91.58%.