METHODS

Datasets

The dataset used in this study was sourced from Kaggle (https://www.kaggle.com/code/mob2dr/98-7-covid-pneumonia-normal-using-xception/data) following a recent revision. After filtration, the final dataset consists of 8,470 Chest X-ray (CXR) images, categorized into three labels: COVID-19, Viral Pneumonia, and Normal (see Table 1). The input images from the dataset were initially converted into matrix format, and the matrices were labeled based on the relative differences observed in the images. The model learns to recognize these differences during the training phase and uses this information to make predictions on the testing dataset.

During the data preprocessing step, all CXR images were resized to 224 x 224 pixels prior to training to address issues such as embedded textual information commonly found in CXR images. Additionally, we retrospectively tested the model on 500 CXR images from our radiology center to assess its accuracy and precision

Feature Extraction and Classification of Chest X-rays

Convolutional Neural Networks (CNNs) leverage spatial and structural information from images. They are capable of learning hierarchical features from data through a backpropagation algorithm. Deep learning (DL) models utilize convolutional layers to extract features from images, which are then used for classification based on these features. The initial layers of DL models typically capture basic attributes such as edges and contours, while deeper layers identify more complex and detailed features of the images (see Fig. 2). This hierarchical learning of features enables the model to progressively refine its understanding of image content, which is crucial for accurate classification.

Proposed System Steps

• Read chest X-ray images
• Image augmentation, enhancement, scaling
• Pre-processing images:

1. Apply image enhancement algorithms to improve image quality.
2. Resize images to 224x224x3 pixels to standardize the input size for the model.
3. Normalize pixel values to a range between 0 and 1 to improve model performance and convergence.
4. Split datasets into training and testing sets (commonly 80% for training and 20% for testing).
5. Extract features.
6. Optimize model performance (Freeze layers, epochs, learning weights, batch size)
7. Evaluate model performance

BACKGROUND

Chest X-ray (CXR) is widely regarded as one of the most effective and reliable methods for diagnosing pneumonia globally (see Fig. 1). It is a rapid, cost-effective, and widely accessible clinical tool that exposes patients to less radiation compared to computed tomography (CT) scans, making it especially valuable in emergency settings and for frequent use in clinical practice (1).  CXR remains a critical diagnostic tool, particularly in resource-limited environments where the availability of advanced imaging modalities such as CT may be limited (2). Developing an accurate, fast, and cost-effective tool for diagnosing viral pneumonia is essential, as it provides the foundation for implementing timely preventive measures such as isolation, contact tracing, and appropriate treatment to curb the spread of infectious diseases like COVID-19 (3). The early detection of pneumonia and COVID-19 through imaging can significantly reduce morbidity and mortality, emphasizing the need to optimize diagnostic techniques.

In recent years, deep learning (DL) approaches have revolutionized the field of medical image classification, providing highly accurate tools for analysing medical images and aiding in the diagnosis of various conditions. These advancements have enabled DL models to perform a wide range of medical image analysis tasks, including detection, staging, and the characterization of pathological abnormalities (4). Specifically, Convolutional Neural Networks (CNNs), a type of deep learning model, have demonstrated exceptional success in image analysis tasks, particularly in medical image diagnostics (5). CNNs have shown great promise in identifying complex patterns in images, enabling them to distinguish between different disease conditions with high precision. Moreover, the use of deep learning models helps minimize human errors in diagnostic processes, leading to more consistent and reliable diagnoses.