Expected Risk Minimization and Robust Preventive Inference of Transfer Learning for COVID-19 Diagnosis within Chest X-Rays

Ahmed, Muhammad A.O.; AbdelSattar, Yasser; Abbas, Ibrahim

doi:10.21608/sjsci.2022.160309.1031

Expected Risk Minimization and Robust Preventive Inference of Transfer Learning for COVID-19 Diagnosis within Chest X-Rays

Document Type : Regular Articles

Authors

¹ Computer Science Department, Faculty of Computers and Information, Luxor University, Luxor 85951, Egypt.

² Mathematics Department, Faculty of Science, Sohag University, Sohag 82524, Egypt

10.21608/sjsci.2022.160309.1031

Abstract

The creation of a treatment strategy and the choice of patient-checking circumstances within many others are supported by early diagnosis of COVID-19 infection. It is possible to detect COVID-19 early on by applying a deep learning method to radiographic medical lab images. Convolutional neural networks (CNN) are used in this study to improve COVID-19 diagnoses using X-ray scans. An automated diagnostic solution that can swiftly deliver accurate diagnostic results is required. CNNs have been found to be efficient at classifying medical images using deep learning techniques. Transfer Learning (TF) is the most reliable research supervised learning method, offering useful analysis to examine many radiographs image samples, and can considerably detect potential and infer preventative detection of COVID-19. Despite its high True Positive, testing healthcare professionals remains a serious risk. Three distinct deep TF and regularization-based architectures were studied on chest X-ray images for the diagnosis of COVID-19. Because these models already include weights trained on the ImageNet database, large training sets are unnecessary. To evaluate the model's performance, 21,165 chest x-ray scan samples were obtained from various sources and identified as COVID-19 data collection from four classes in the Kaggle repository. Average metrics results are collected to get the actual predictions for all classes. Although Saving training time with TF, an advance improvement for performance can be achieved by applying only some parts of the input image with most important segments of the input image are localized. To prove the validity of our approach we use Grad Cam algorithm to find the input image parts with most valuable features for decision making. The localised image region map is udsed to reproduce a lighter version of the image database with only marked as most important image regions. Metrics including precision, F1-Score, confusion matrix, accuracy, sensitivity, specificity, error rate, and error rate have been used to assess the performance of all the TF models., besides false positive (FP), Matthews Correlation Coefficient (MCC), and Kappa performance measures. In terms of performance, the ResNet-50 model outperforms all others with a low error rate of 0.039 and achieves more than a 96% accuracy. The study findings proven the proposed model validity as a computer-aided diagnostics model with a guarantee to supply help for radiologists quickly and accurately.

Keywords