Detection of pulmonary diseases from chest X-ray images using deep learning model

Abstract

Deep learning models are important in efficiently identifying different pulmonary diseases from Chest X-ray Images (CXRs). Pneumonia is one of the most common lung diseases that cause death. Especially, stage 4 pneumonia can become the reason for an untimely death. Moreover, COVID-19 is still killing a lot of people all around the world. Scientists, doctors, and institutions are working on inventing the most effective way of detecting these diseases. Accurate and early detection of these diseases is essential, otherwise, they can be deadly. In this work, we will detect different pulmonary diseases like COVID-19, and Pneumonia from chest X-ray images. There are many deep learning models like CNNs, RNNs, GANs, and so on. Among them, CNN models are the best for image classification. For example, ResNet18, ResNet50, InceptionV3, VGG19, DenseNet201 and so on. However, we have not used these models. We have used models that have the highest accuracy, Recall, precision, and F1 score. The CNN models generally perform well with image data. So, we used models that are not traditional CNN models. Rather, they essentially rely on transformer architectures or a combination of transformers and CNNs. So, we have used a Swin Transformer, Vision Transformer (ViT), VoLO-D1, FocalNet, and VITamin. Transformers rely on self-attention mechanisms to determine the similarities across an image. On the other hand, CNNs use convolutional layers to extract features locally from an image. Our proposed model is a customized CNN model and it is time and cost-efficient as it provides higher accuracy faster than other models. It is deploy-friendly as the size of the model is 257 MB. Other transformer based model are bigger in size. Moreover, it has a transformer-based ecosystem and benefits. The accuracy of our customized CNN model is 98 percent and learning rate is 0.001. We have built an automated lung disease detection system to make the detection less time-consuming, cost-efficient, and error-free for developing countries.