딥러닝 기반 위 질환 컴퓨터 보조 진단 시스템 개발 및 Grad-CAM을 활용한 시각화 (Development of Computer-aided Diagnosis System for Gastric Disease based on Deep Learning and Visualization using Grad-CAM)

서지정보

· 발행기관 : 대한전기학회
· 수록지 정보 : 전기학회논문지 / 72권 / 2호 / 234 ~ 240페이지
· 저자명 : 이한성, 조현종

초록

Gastric cancer ranked first in the incidence of cancer in Korea from 2015 to 2018, and it is a trend that many people suffer from regardless of gender. Because the cause of gastric cancer is not clear and early gastric cancer has no specific symptoms, patients often miss the appropriate treatment time. Gastric endoscopy is the most effective method for diagnosing gastric cancer, but early gastric cancer accompany only subtle changes that are difficult to distinguish with the naked eye, and the diagnosis accuracy may vary depending on the proficiency of the specialist. Therefore, we proposed an EfficientNetV2-L-based Computer-assisted Diagnosis (CADx) system for gastric diseases in this paper. EfficientNetV2 is a network that can have high classification performance for gastric diseases even with a small number of parameters as a model designed using the Training-NAS method. In general, the diversity of data affects the performance of deep learning models. So, we augmented data using Cifar10 augmented policy. Through this method, sensitivity was improved from 0.878 to 0.966 in early gastric cancer and normal classification, and from 0.690 to 0.817 in abnormal and normal classification. In addition, we tried to visualize the region of interest of the model for the location of the lesion by applying Grad-CAM.

영어초록

Gastric cancer ranked first in the incidence of cancer in Korea from 2015 to 2018, and it is a trend that many people suffer from regardless of gender. Because the cause of gastric cancer is not clear and early gastric cancer has no specific symptoms, patients often miss the appropriate treatment time. Gastric endoscopy is the most effective method for diagnosing gastric cancer, but early gastric cancer accompany only subtle changes that are difficult to distinguish with the naked eye, and the diagnosis accuracy may vary depending on the proficiency of the specialist. Therefore, we proposed an EfficientNetV2-L-based Computer-assisted Diagnosis (CADx) system for gastric diseases in this paper. EfficientNetV2 is a network that can have high classification performance for gastric diseases even with a small number of parameters as a model designed using the Training-NAS method. In general, the diversity of data affects the performance of deep learning models. So, we augmented data using Cifar10 augmented policy. Through this method, sensitivity was improved from 0.878 to 0.966 in early gastric cancer and normal classification, and from 0.690 to 0.817 in abnormal and normal classification. In addition, we tried to visualize the region of interest of the model for the location of the lesion by applying Grad-CAM.

참고자료

· 없음