<![CDATA[Gastroesophageal Reflux Disease (GERD) is a common digestive tract disorder, and endoscopic image-based diagnosis still faces challenges due to variations in lighting, image capture angles, and limited data that can cause deep learning models to overfit and reduce generalization capabilities. While previous studies frequently prioritize overall accuracy, this study distinguishes itself by focusing on increasing recall as a critical indicator of clinical sensitivity for positive case detection. This study aims to analyze the effect of applying image augmentation on the performance of GERD endoscopic image classification using a lightweight MobileNetV2 architecture, offering a practical solution for resource-constrained clinical settings. The research methods include data collection, class mapping, 80:20 data split, and standardized preprocessing stages. Two training scenarios were compared: without augmentation and with strategic on-the-fly augmentation applied through a transfer learning approach. Evaluation was carried out using accuracy, precision, recall, F1-score, and ROC-AUC. The results showed that augmentation improved model performance, with accuracy increasing from 76.72% to 78.06%, recall increasing from 79.33% to 85.27%, F1-score increasing from 77.86% to 80.04%, and AUC increasing from 0.8317 to 0.8481. These findings indicate that the augmentation strategy effectively improves the generalization of the lightweight model, significantly reducing missed diagnoses to support early detection of GERD. In addition, the results demonstrate that applying augmentation techniques can help the model learn more robust visual features from limited medical image datasets. This approach also contributes to reducing overfitting and improving model reliability in practical clinical image analysis scenarios.]]>
Copyrights © 2026
