Pneumonia Detection with Dual-Backbone Feature Fusion using Deep Learning

Abstract

Chest X-ray automated pneumonia detection is an important research field in medical image analysis. While DCNNs have been successful on benchmark datasets, many models fail to perform as well over different datasets when there is domain shift. To this end, this paper proposes a cross-dataset pneumonia classification method, which combines dual-backbone feature fusion and lightweight domain adaptation. The proposed method uses two parallel feature extractors (EfficientNet-B0 and MobileNetV3-Small) and then uses projection layers, interaction-based fusion of parallel feature extractors, and an adaptive feature selection mechanism through a gating mechanism. Correlation alignment (CORAL) loss and mean feature alignment loss are used in training to minimize the difference between the source domain and the target domain. The dataset used for training the model is RSNA Pneumonia dataset, and the target dataset is the COVID-19 Radiography dataset. Experimental results show that the proposed method outperforms the other methods significantly in terms of target-domain accuracy (91.62%) and AUC (95.41%), which indicates that the proposed method has good cross-dataset generalization performance. The results demonstrate that the use of complementary backbone representations along with domain-aware alignment is an effective approach to achieving robust medical image classification.