Viral Pneumonia and Covid Screening on Lung Ultrasound

Abstract

The rise of Covid-19 pandemic has exaggerated the necessity for safe, quick and sensitive diagnostic tools to confirm the protection of tending employees and patients. Although ML has shown success in medical imaging, existing studies concentrate on Covid-19 medicine victimization using Deep Learning (DL) with X-ray and computed axial Tomography (CT) scans. During this study we tend to aim to implement CNN model on Lung Ultrasound (LUS), to assist doctors with the designation of Covid-19 patients. We selected LUS since it's quicker, cheaper and additional out there in rural areas compared to CT and X- ray. We have used the biggest public dataset containing LUS pictures and videos of Covid, Pneumonia and healthy patients that has been collected from totally different resources. We tried out frame level approach that extracted 5 frames per patient video. We'll use this dataset to experiment with a CNN model that has hyper parameter calibration. We conjointly enclosed explainable AI using Grad-CAM that uses gradients of a selected target that flows through the convolutional network to localize and highlight regions of the target within the image. Moreover, we'll experiment with completely different data preprocessing techniques that may aid with pattern recognition and increasing the DL model’s accuracy like histogram equalization, standardization, Principle Component Analysis (PCA) and Synthetic Minority Oversampling Technique (SMOTE). Lastly, we tend to create a straightforward application that diagnoses LUS videos with our CNN model, and shows the frame results with visual illustration of why the model has taken certain prediction with the help of Gradient-Weighted category Activation Mapping (Grad-CAM).