Chest X-Ray Image to Classify Lung diseases in Different Resolution Size using DenseNet-121 Architectures
Ovy Rochmawanti, Fitri Utaminingrum
Abstract
Chest radiography (CXR) is the most commonly used diagnostic tool in medical practice because of the low cost and easy operation. CXR contains much information about a patient's health, and radiologists often use it for disease detection. The diagnoses are often subjective for a few reasons, like illness looks, which might be unclear in CXR images or are often confused with different diseases. In this study, DenseNet-121 is a well-known convolutional neural network (CNN) model for diagnosing illness. The convolutional layers of these models are used as a base network. The pre-trained model is used because this study applied the transfer learning technique. The pre-trained model is built and trained using the public domain dataset ImageNet.Global Average Pooling (GAP) and dropout layers are added to reduce the overfitting problem of the network. The batch normalization layer is used for the rapid training of the pre-trained model. The output layer consists of 2 nodes that directly represent the two classes and a softmax activation function.This study analyzes the effects of varying image resolution for CXR images using four different datasets: tuberculosis dataset, pneumonia dataset, cardiomegaly dataset, and COVID-19 dataset. It is experimentally shown that DenseNet121 model achieves the highest accuracy in classification using image size 224x224 pixels. The best results were obtained with Tuberculosis dataset, Pneumonia dataset, Cardiomegaly dataset and COVID-19 dataset with 0,892, 0,904, 0,898 and 0,986, respectively.