Enhanced Classification of Gastric Lesions and Early Gastric Cancer Diagnosis in Gastroscopy Using Multi-Filter AutoAugment

Abstract

Gastric cancer is high-risk cancer in terms of both incidence and mortality. However, if it is diagnosed early, there is a high chance of survival. Therefore, an early diagnosis of gastric cancer and precancerous lesions is very important. Gastroscopy is one of the best methods for diagnosing gastric cancer and precancerous lesions, but it relies on visual observation by medical specialists. Accordingly, factors such as the experience or fatigue of specialists can influence diagnosis results. To alleviate these problems, we propose a computer-aided diagnosis system that can improve the efficiency of diagnosis and reduce misdiagnoses by providing a second opinion. We aimed to classify healthy tissue, gastric lesions, and early gastric cancer using a Transformer-based deep-learning classification model called Vision Transformer, which has achieved the best performance in transfer learning. We also proposed a Multi-Filter AutoAugment (MFAA) method, which increases the classification performance of the model given small amounts of medical data. The medical data augmented using MFAA are better for training deep-learning models than conventionally augmented data; we effectively enhanced the classification performance of the model using MFAA. In experiments, the model achieved an F1-score of 0.87 and area under the curve of 0.94 in the classification of abnormalities (gastric lesions including early gastric cancer) and healthy tissue. In addition, it obtained an F1-score of 0.92 and area under the curve of 0.97 in the classification of early gastric cancer and non-cancerous gastric lesions.