CX-Net: Multi-scale Enhanced Fusion Network for Pneumonia Classification

Due to the blurred nature of lung X-ray images and the variability in lesion density and lesion location, even relatively experienced radiologists can have diagnostic difficulties. Deep Learning (DL) can quickly and accurately determine the type of pneumonia from an X-ray. However, the methods proposed so far will extract many invalid features during extraction, and the pneumonia lesions are location-specific, and the features around the lesions often affect the final classification accuracy. Therefore, this paper proposes a multi-scale local enhancement model (CX-Net) consisting of a multi-scale fusion module, a local channel attention module, and a lung enhancement module to solve this problem. The local channel attention module (LCAM) performs local data evaluation on the features extracted by the encoding network ResNet-50 to assign weights to the feature layer. The Lung Enhancement Module (LEM) enhances the data of possible diseased areas to highlight the specific features of the lung lesion location. The enhanced features are multi-scale superimposed to obtain a matrix with attentional bias, thus avoiding the influence of useless features to a greater extent. In this paper, different public data sets in Kaggle are merged into a binary classification dataset, a three-classification dataset, and a four-classification dataset to test the performance of the model. Finally, it is shown through extensive experiments that the results of the proposed model in this paper outperform the current state-of-the-art methods. The models proposed in the two classifications, three classifications, and four classifications achieved a classification accuracy of 98.34%, 97.62%, and 94.87%.