Pneumonia Detection Using Transfer Learning And Hardware Implementation in Edge TPU

In the modern-day scenario, Machine Learning (ML) holds an important role in providing automation, dataset-based decision-making, security, and healthcare advancements. This paper presents an idea of how Machine Learning algorithms can be used to detect a deadly disease like Pneumonia. It is a problem of image classification where various Convolutional Neural Network (CNN) architectures are used to classify whether the given chest X-ray has Pneumonia or not. At first, a dataset having 5856 chest X-ray images is collected and some data pre-processing is done. These include resizing images, pixel value normalization, and image augmentation (horizontal flip, rotation, etc). Four models are created, one by using CNN architecture and the other three by Transfer Learning using pre-trained models of standard architectures (Resnet50, Densenet121 and VGG19). All four models are trained with same dataset and a comparative study is presented in terms of architectures and performance. With the CNN model, highest accuracy of 92.6% is obtained. Models made using pre-trained models of Resnet50, Densenet121 and VGG19 gave accuracies of 77%, 91% and 91% respectively. This paper also proposes hardware implementation of the CNN model in the Coral Dev board with Edge TPU. After implementation, it is able to classify whether the given sample chest X-ray is ‘NORMAL’ or ‘PNEUMONIA’.