ABSTRACT
The COVID-19 triggered a worldwide public health dilemma. The number of COVID-affected persons and their demise was increasing on a daily basis, putting enormous strain on our social and healthcare systems. Because the disease is very contagious, the only approach to limit its spread is to separate affected people from non-infected people. As a result, rapid detection of COVID-19 instances is an important step in combating this virus and thus protecting the healthcare system. A lengthy clinical testing period is one of the main elements contributing to the rapid spread of the COVID-19 pandemic. An imaging tool, including a chest X-ray (CXR), can assist in speeding up the identification procedure. As a result, utilizing CXR pictures, an automatic CAD system is built to recognize COVID-19 samples from healthy people and respiratory disease patients. A total of 300 COVID chest X-ray images having a posteroanterior (PA) view were taken into account from dataset to be fed to the CNN algorithm. The lungs, bony thoracic cavity, mediastinum, and major vessels are all evaluated in the PA chest view. Of which, 70% of images are used for training and the remaining 30% is split between 15% for testing and 15% for validation. Data is processed using Pandas and NumPy, and necessary Image DataGenerator parameters are applied to make the model robust and effective to distorted images. NumPy arrays are very fast and require enormously less memory and Pandas is very effective for data manipulation-related stuff, which is why it is considered for overall effectiveness in model making. Thereafter, initially, data is fed to the CNN model with 11 layers comprehensive of ConV2d, max pooling, dropout and dense layers with 20 epochs. The model achieved the peak accuracy of 91% initially. Upon hyperparameter tuning, the model achieved the peak accuracy of 96.88%. A spike of 5.88% in accuracy is a consequence of adding six new layers and also hyperparameter tuning. The trained model is saved in form of .h5 file. Android supports TensorFlow Lite files to integrate with. As a consequence, .h5 file is converted into .tflite file using necessary operations. In this way, the model is successfully integrated into the Android application.
Keywords: COVID-19, CNN
