Utilizing Incremental Learning-Based EfficientNet for Chest X-Ray Classification

Chest X-ray image classification, especially for pneumonia detection, faces challenges in incremental learning, such as accuracy degradation and catastrophic forgetting. This study proposes a method combining EfficientNet-B0 with Analytic Class Incremental Learning (ACIL) to address these issues. EfficientNet-B0 enhances feature extraction through compound scaling, while ACIL mitigates forgetting by updating an autocorrelation matrix without storing historical data. Comparative experiments with ResNet-18 and three incremental learning methods (LwF, ER, EWC) show that EfficientNet-B0 with ACIL achieves superior performance. In binary classification, it reaches 78.3% accuracy with a forgetting rate of −7.3%. In multi-class classification, accuracy reaches 68.1%, with only 1.0% forgetting. These results demonstrate the proposed method’s effectiveness and stability for incremental learning in chest X-ray image analysis.