Comparative Performance Analysis of Convolutional Neural Network and EfficientNet-B0 Transfer Learning for Pneumonia Detection in Chest X-Ray Images

Abstract

Pneumonia is an acute lung infection and remains one of the leading causes of mortality among children under five in Indonesia. Pneumonia is commonly diagnosed through the analysis of chest X-ray (CXR) images by radiologists; however, this process is still subject to human variability and the limited availability of medical specialists. This study aims to develop a Convolutional Neural Network (CNN)-based pneumonia detection system by comparing four model variants: CNN with Color Jitter (Model 1), CNN with Color Jitter and Layer Augmentation (Model 2), CNN with Batch Normalization and Color Jitter (Model 3), and EfficientNet-B0 transfer learning (Model 4). The experiments were conducted using the Chest X-Ray Images (Pneumonia) dataset from Kaggle, comprising 5,856 images, which were divided into 70% for training, 15% for validation, and 15% for testing. The experimental results demonstrate that the EfficientNet-B0 transfer learning model achieved the best performance, with an accuracy of 96.92%, precision of 97.67%, recall of 98.13%, and an F1-score of 97.90%. This model improved classification accuracy by 1.93% compared with the conventional CNN baseline (Model 1), which achieved an accuracy of 94.99%. Furthermore, the incorporation of Layer Augmentation in Model 2 effectively reduced the number of false negatives from 21 to 10 cases. These findings demonstrate the effectiveness of EfficientNet-B0 transfer learning for medical image classification and highlight its potential as a clinical decision support tool for pneumonia diagnosis in healthcare settings, rather than as a replacement for clinical diagnosis performed by qualified medical professionals.