A comprehensive fusion model for improved pneumonia prediction based on KNN-wavelet-GLCM and a residual network

Abstract

Pneumonia is a severe disease that contributes to global mortality rates, emphasizing the critical need for early detection to improve patient survival. Chest radiography (X-ray) images serve as a fundamental diagnostic tool in clinical practice to detect various lung abnormalities. However, medical images, particularly X-rays, contain crucial data that are often imperceptible to the human eye. This study presents a novel fusion model (Res-WG-KNN) based on a soft voting ensemble strategy to predict pneumonia from chest X-ray images. It utilizes 2D-discrete wavelet decomposition and texture features from the Gray Level Co-occurrence Matrix (GLCM) with supervised machine learning, alongside raw X-ray images using a modified Residual Network ResNet-50. The proposed model was evaluated using two public pneumonia X-ray image datasets: one for adult patients, called the Radiological Society of North America (RSNA) dataset, and one for pediatric patients, called the Kermany dataset. These datasets differ in both size and image format, with the RSNA dataset using DICOM images and the Kermany dataset using JPEG images. The use of a soft voting technique in the proposed model effectively enhances classification performance beyond current benchmarks, achieving 97.0% accuracy and 0.97 AUC on the RSNA dataset, and 99.0% accuracy with 0.99 AUC on the Kermany dataset for pneumonia prediction.