Abstract:
The COVID-19 outbreak made it evident how crucial it is to have diagnostic tests that are both accurate , quick and cheap. RT-PCR is the usual test that many people use, however it often takes a long time to get results and can sometimes give false negatives . These limitations prompted us to investigate the potential of deep learning as a more expedient and efficient method for detecting COVID-19 directly from chest X-ray (CXR) pictures. In this study, we created a balanced dataset of 3,100 X-ray images—1,575 confirmed COVID19 cases and 1,525 normal cases. Some of the photographs were from Shahjalal Medical Hospital, and some came from open-access archives. We did certain preprocessing measures before training, such scaling, normalizing, and adding more data, to make the images more homogeneous and help the models generalize better.We examined both a lightweight proprietary Convolutional Neural Network (CNN) and five well-known pretrained architectures: VGG16, ResNet50, DenseNet121, InceptionV3, and Xception to see how well they worked. All of the tests were done on Google Colab, which has GPU support. We used common assessment measures like accuracy, precision, recall, F1-score, and ROCAUC, and we used categorical cross-entropy as the loss function and Adam as the optimizer.The results were very surprising. VGG16 had the best accuracy at 97% and the same great F1-score. Our bespoke CNN, which is significantly simpler, nevertheless got 92%, which makes it useful in real-world settings like hospitals with limited resources and few computers. Once again, our results show that transfer learning can greatly increase the accuracy of medical image classification. The results also demonstrate that smaller, lighter models can be useful when speed and efficiency are important .In general, this experiment shows that deep learning has a lot of potential for helping doctors diagnose COVID-19 faster. More importantly, it makes it possible to use similar methods on other respiratory
