Comparative Study of Deep Learning Architectures for Pneumonia Diagnosis from Chest X-Ray Images

This paper presents an end-to-end pneumonia diagnosing method from CXR images. We investigate the effectiveness of feature extraction using deep learning models, namely ResNet50, ViT, and Inception v3, besides the performance of Random Forest and Logistic Regression classifiers stemming from these models. An in-depth performance analysis was done in terms of accuracy, F1 score, and AUC values to determine which model works best. The research work discussed herein focuses on the challenge of pneumonia diagnosis manually and determines the urgent need for automated solutions, concluding that a combination of Inception v3 and Logistic Regression achieved high accuracy of 0.7948, high F1-score of 0.6279, and ideal Kappa Coefficient value of 0.5622. This enhances the clinical efficiency and accuracy in pneumonia detection that will be helpful in patients' treatment outcomes. The dataset utilized in this study consists of pediatric chest X-ray images from Guangzhou Women and Children's Medical Center, with rigorous quality control and expert review to ensure diagnostic accuracy for AI training.