A Comparative Analysis of Chest X-Rays and CT Scans Towards COVID-19 Detection

Abstract

The diagnosis of COVID-19 is greatly aided by radiological chest examinations. In this study, we conduct a comprehensive comparison of various classification models to classify COVID-19 from Chest X-ray images and devise an automated, precise model to segment the COVID-19 infected regions present in Chest CT Scans and make the results interpretable with the use of GradCAM. A total of 4000 images from the COVID-19 Radiography dataset were used for training and evaluating the models. In order to predict COVID-19 in Chest X-rays, the transfuse segmentation model [1] is used to segment the lungs. Radiomic features are then extracted and the most essential features were selected using Principle Component Analysis which were then used to train six classification models, namely support vector machines (SVM), Logistic Regression (LR), Decision Trees, AdaBoost (AB), K-Nearest Neighbors (KNN) and Random Forest. Support vector machines was the best-performing classifier with an accuracy of 0.9115 and an AUC of 0.9117. In the case of CT scans, lungs were segmented using image processing techniques, and a total of 2482 CT Scans from the SARS-COV-2 CT-Scan Dataset were used to train a ResNet18 CNN architecture, whose results were made interpretable by employing gradCAM. Furthermore, infections were segmented from the CT scans using two image processing based approaches, the first making use of the Discrete Wavelet Transform and the second approach performing morphological operations. The ResNet18 CNN used for assisting GradCAM achieved a testing accuracy of 96.34%. In order to evaluate the results, the infections were compared with 25 manually segmented CT scans and achieved a precision of 0.975, a recall of 0.960, and an IOU score of 0.937.

Ähnliche Arbeiten