A Hybrid CNN with Feature Fusion for Patch-Level Interstitial Lung Disease Classification

Abstract

The classification of Interstitial Lung Disease (ILD) patterns in Computed Tomography (CT) images is challenging due to the irregular patterns of the disease and the similarity between patterns of different tissues. Moreover, the performance of classification models can be degraded by severe class imbalance in the dataset. In this study, we propose a hybrid Convolutional Neural Network (CNN) with feature fusion for patch-level ILD classification. The proposed framework effectively captures texture features from raw images while also considering contextual features at multiple scales. The class imbalance problem is addressed using the Balanced SoftMax loss function. Additionally, the proposed framework is trained using the Stochastic Weight Averaging (SWA) technique to improve model stability. The framework is trained and tested using CT image patches from five different ILD classes. The proposed framework achieves high performance, with accuracies of 98.23% on the validation dataset and 98.09% on the test dataset.