DI-EffNet: A Dual-Attention Network for Binary ILD Classification from Imbalanced CT Data

Abstract

Classifying computed tomography (CT) images for interstitial lung diseases (ILDs) is a significant research challenge. Recent studies have explored the effectiveness of pre-trained models to improve performance and accuracy. However, training on imbalanced datasets remains a major hurdle, often resulting in biased predictions. This study focuses on multi-label classification of thoracic lung diseases using CT images, specifically addressing class imbalance. To mitigate this issue, data augmentation methods were employed to create synthetic samples, and images from various sources were integrated, improving the models’ generalization capabilities. In this paper, we introduce DI-EffNet, a novel dual-input neural network architecture that combines both local and global attention mechanisms to enhance the binary classification of ILDs on CT scans. Comparative experiments show that DI-EffNet significantly outperforms established deep learning models, achieving an accuracy of 99.75%, compared to 89.39% for ResNet-50, 93.9% for VGG-19, and 93.29% for EfficientNet B0. These results demonstrate that DI-EffNet provides a robust and effective solution for ILD detection, with strong potential to support clinical diagnosis.