An Enhancement to Fused and Cascaded Squeeze and Excitation Network for Pneumonia Prediction with Deep Learning Techniques

Abstract

Pneumonia is a rapidly spreading disease that poses a significant risk to the health and welfare of those affected. Accurately diagnosing pneumonia from a biomedical perspective requires the use of various diagnostic tools and the assessment of multiple clinical characteristics. However, the lack of available experts and equipment hinders the proper diagnosis of this disease. As deep learning is getting revolutionized in terms of these fields, the utilization of such advanced models brings more efficacy. This study aims to comprehensively enhance both the Inception V3-based squeeze-excitation network and the Densenet 121 with squeeze-excitation network and analyze the performance of both in addition to state-of-the-art models, for accurate pneumonia detection from medical images. This paper aims to develop a robust deep-learning model capable of effectively identifying Pneumonia by analyzing the characteristics derived from chest X-ray images. The study seeks to identify the most efficient model for pneumonia detection by comparing and assessing the performance of these models. The study consists of three main phases. a) Data collection from Kermany and RSNA Pneumonia challenge datasets, which consist of X-ray images, b) Preprocessing the data to remove noises and anomalies using a Gaussian filter to improve image quality, and c) Conducting the classification task by utilizing the Densenet 121-based squeeze-excitation network and the Inception V3-based squeeze-excitation network. The Inception V3 SENet model demonstrated superior performance. With an accuracy of 0.9539%, recall of 0.95%, and f1-score of 0.95%, it surpassed the Densenet model's performance in terms of accuracy (0.74), recall (0.81), and f1-score (0.69), as well as outperformed other state-of-the-art models.