SAM Optimizer를 통한 위내시경 이미지 분류 CADx의 성능 향상 연구

Abstract

Gastric cancer has a high incidence in East Asians, and the risk increases over time. Often, gastric cancer presents no early symptoms, leading to missed treatments. Consequently, in Korea, support is provided to individuals over 40 years of age who undergo gastroscopy. However, as the number of gastroscopy patients increases, doctors' fatigue rises, becoming a factor that can lead to misdiagnosis. Therefore, this paper proposes a CADx (Computer-Aided Diagnosis) system for gastric lesion classification based on ConvNeXt and ViT (Vision Transformer), applying the SAM (Sharpness Aware Minimization) optimizer. ConvNeXt is a network that achieves high performance by incorporating techniques from Swin Transformer and the latest advancements, with ResNet-50 as the base model. ViT divides the image into smaller patches and uses these patches as input to the Transformer. This allows for learning relationships between patches and ultimately leads to image classification. To address the issue of limited data in medical images, the gastric abnormal dataset was augmented using the AutoAugment policy. The SAM Optimizer is an optimization technique that detects and minimizes the "sharpness" of the loss function that may occur during the deep learning model's learning process. Using this method, the sensitivity of classifying abnormal and normal gastroscopy images in ConvNeXt increased from 0.7167 to 0.9583 for the original dataset and from 0.7583 to 0.9833 for the augmented dataset. ViT exhibited a significant decrease from 0.9500 to 0.7750 in the original dataset but increased from 0.9500 to 0.9583 in the augmented dataset. This demonstrates that the SAM Optimizer can effectively enhance CADx performance.