COVID-19 diagnosis: A comprehensive review of pre-trained deep learning models based on feature extraction algorithm

Area of Science:

• Medical Imaging and Artificial Intelligence
• Deep Learning in Diagnostics
• Radiography Analysis

Background:

• The rise of COVID-19 necessitates rapid and accurate diagnostic methods.
• Traditional RT-PCR testing is time-consuming and expensive.
• Medical imaging, specifically chest X-rays, offers a feasible alternative for COVID-19 diagnosis.

Purpose of the Study:

• To investigate the potential of Artificial Intelligence (AI)-based early diagnosis of COVID-19 using X-ray chest radiographs.
• To optimize deep transfer learning models for accurate COVID-19 detection.
• To compare the performance of various classifiers for COVID-19 diagnosis from radiographic images.

Main Methods:

• Utilized a dataset of 10,192 normal and 3616 COVID-19 chest X-rays.
• Applied data augmentation to enhance the training dataset.
• Trained multiple deep transfer learning models (including Inception-V3) and evaluated them using classifiers like SVM, KNN, NN, Naive Bayes, and Logistic Regression.
• Assessed model performance using metrics such as accuracy, precision, F1 score, recall, and AUC.

Main Results:

• Inception-V3 achieved a training accuracy of 84.79% among the deep transfer learning models.
• Support Vector Machine (SVM) classifiers, particularly Cubic SVM, showed superior performance with an AUC of 0.99, accuracy of 95.8%, precision of 0.983, recall of 0.8977, and F1 score of 0.9384.
• SVM classifiers outperformed other tested classifiers (KNN, NN, Naive Bayes, Logistic Regression) across key performance metrics.

Conclusions:

• AI-based analysis of chest X-rays using deep transfer learning models shows significant potential for early COVID-19 diagnosis.
• SVM classifiers are highly effective for classifying COVID-19 from X-ray images, providing high accuracy and reliability.
• The proposed methodology offers a valuable tool for preliminary COVID-19 diagnosis, complementing existing methods.