ABSTRACT
Chest x-ray images are used as a primary diagnostic tool for different thoracic diseases like Pneumonia, Covid-19, SARS, Pneumothorax, etc. The abnormalities in these chest x-ray images are sometimes subtle and require expert eyes to identify. In this paper, we proposed a deep learning-based tool that automatically detects abnormalities in different chest radiography images and is able to identify with the highest probability if the patient is suffering from Covid- 19, Effusions, Infiltration, Pneumonia, or Pneumothorax. Convolutional Neural Network (CNN) based architecture, named ChestXRNet, supervised multi-class classification technique, is proposed to classify different thoracic chest x-ray images. The training dataset was compiled by collecting samples from three different open-source databases. We only scraped the necessary data for this project and built a customize dataset consists of 16,200 images. The performance of the proposed ChestXRNet model is compared with some pre-trained CNN models like DenseNet201, EfficientNetB7, and VGG16 for benchmarking using the same dataset. Our model acquired a training evaluation accuracy of 92.9% and a testing evaluation (prediction) accuracy of 82.8%. Model weights were saved and then integrated into a Flask based web server for a proper web application experience.
