Résumé
COVID-19 has been raging for almost three years ever since its first outbreak. It is without a doubt that it is a common human goal to end the pandemic and how it was before it started. Many efforts have been made to work toward this goal. In computer vision, works have been done to aid medical professionals into faster and more effective procedures when dealing with the disease. For example, disease diagnosis and severity prediction using chest imaging. At the same time, vision transformer is introduced and quickly stormed its way into one of the best deep learning models ever developed due to its ability to achieve good performance while being resources friendly. In this study, we investigated the performance of ViT on COVID19 severity classification using an open-source CXR images dataset. We applied different augmentation and transformation techniques to the dataset to see ViT's ability to learn the features of the different severity levels of the disease. It is concluded that training ViT using the horizontally flipped images added to the original dataset gives the best overall accuracy of 0.862. To achieve explainability, we have also applied Grad-CAM to the best performing model to make sure it is looking at relevant region of the CXR image upon predicting the class label. © 2022 IEEE.
Résumé
COVID-19 and lung diseases have been the major focus of research currently due to the pandemic's reach and effect. Deep Learning (DL) is playing a large role today in various fields from disease classification to drug response identification. The conventional DL method used for images is the Convolutional Neural Network (CNN). A potential method that will replace the usage of CNNs is Transformer specifically Vision Transformers (ViT). This study is a preliminary exploration to determine the performance of using ViT on diseased lungs, COVID-19 infected lungs, and normal lungs. This study was performed on two datasets. The first dataset was a publicly accessible dataset from Iran that has a large cohort of patients. The second dataset was a Malaysian dataset. These images were utilized to verify the usage of ViT and its effectiveness. Images were segregated into several sized patches (16x16, 32x32, 64x64, 128x128, 256x256) pixels. To determine the performance of ViT method, performance metrics of accuracy, sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV) and F1-score. From the results of this study, ViT is a promising method with a peak accuracy of 95.36%. © 2021 IEEE.