Wavelet and deep learning-based detection of SARS-nCoV from thoracic X-ray images for rapid and efficient testing

This paper proposes a wavelet and artificial intelligence-enabled rapid and efficient testing procedure for patients with Severe Acute Respiratory Coronavirus Syndrome (SARS-nCoV) through a deep learning approach from thoracic X-ray images. Presently, the virus infection is diagnosed primarily by a process called the real-time Reverse Transcriptase-Polymerase Chain Reaction (rRT-PCR) based on its genetic prints. This whole procedure takes a substantial amount of time to identify and diagnose the patients infected by the virus. The proposed research uses a wavelet-based convolution neural network architectures to detect SARS-nCoV. CNN is pre-trained on the ImageNet and trained end-to-end using thoracic X-ray images. To execute Discrete Wavelet Transforms (DWT), the available mother wavelet functions from different families, namely Haar, Daubechies, Symlet, Biorthogonal, Coiflet, and Discrete Meyer, were considered. Two-level decomposition via DWT is adopted to extract prominent features peripheral and subpleural ground-glass opacities, often in the lower lobes explicitly from thoracic X-ray images to suppress noise effect, further enhancing the signal to noise ratio. The proposed wavelet-based deep learning models of both, two-class instances (COVID vs. Normal) and four-class instances (COVID-19 vs. PNA bacterial vs. PNA viral vs. Normal) were validated from publicly available databases using k-Fold Cross Validation (k-Fold CV) technique. In addition to these X-ray images, images of recent COVID-19 patients were further used to examine the model's practicality and real-time feasibility in combating the current pandemic situation. It was observed that the Symlet 7 approximation component with two-level manifested the highest test accuracy of 98.87%, followed by Biorthogonal 2.6 with an efficiency of 98.73%. While the test accuracy for Symlet 7 and Biorthogonal 2.6 is high, Haar and Daubechies with two levels have demonstrated excellent validation accuracy on unseen data. It was also observed that the precision, the recall rate, and the dice similarity coefficient for four-class instances were 98%, 98%, and 99%, respectively, using the proposed algorithm.