Cloud Hosting

SARS-CoV-2, known as corona virus, causes COVID-19. It is an infectious disease first discovered in China in December 20191,2,3. World Health Organization (WHO) also declares it as a pandemic. Figure1 shows its detail structure3. This new virus quickly spread throughout the world. Its effect is transmitted to humans through their zoonotic flora. COVID-19's main clinical topographies are cough, sore throat, muscle pain, fever, and shortness of breath4,5. Normally, RT-PCR is used for COVID-19 detection. CT and X-ray have also vital roles in early and quick detection of COVID-196. However, RT-PCR has low sensitivity of about 60% -70% and even some times negative results are obtained7,8. It is observed that CT is a subtle approach to detecting COVID-19, and it may be a best screening means9.

Artificial intelligence and its subsets play a significant role in medicine and have recently expanded their prominence by being used as tool to assist physicians10,11,12. Deep learning techniques are also used with prominent results in many disease detections like skin cancer detection, breast cancer detection, and lung segmentation13,14. However, Due to limited resources and radiologists, providing clinicians to each hospital is a difficult task. Consequently, a need of automatic AI or machine learning methods is required to mitigate the issues. It can also be useful in reducing waiting time and test cost by removing RT-PCR kits. However, thorough pre-processing of CT images is necessary to achieve the best results. Poisson or Impulse noise during the acquisition process of these photos could have seriously damaged the image information15. To make post-processing tasks like object categorization and segmentation easier, it is essential to recover this lost information. Various filtering algorithms have been proposed to de-blur and to de-noise images in past. Standard Median Filter (SMF) is one of the most often used non-linear filters16.

A number of SMF modifications, including Weighted median and Center weighted median (CWM)17,18, have been proposed. The most widely used noise adaptive soft-switching median (NASM) was proposed in19, which achieved optimal results. However, if the noise density exceeds 50%, the quality of the recovered images degradedsignificantly. These methods are all non-adaptive and unable to distinguish between edge pixels, uncorrupted pixels, and corrupted pixels. Recent deep learning idea presented in20,21,22 performs well in recovering the images degraded by fixed value Impulse noise. However, its efficiency decreases with the increase in the noise density and in reduction of Poisson noise, which normally exist in CT images. Additionally, most of these methods are non-adaptive and fails while recovering Poisson noise degraded images. In the first phase of this study, layer discrimination with max/min intensities elimination with adaptive filtering window is proposed, which can handle high density Impulse and Poisson noise corrupted CT images. The proposed method has shown superior performance both visually and statistically.

Different deep learning methods are being utilized to detect COVID-19 automatically. To detect COVID-19 in CT scans, a deep learning model employing the COVIDX-Net model that consists of seven CNN models, was developed. This model has higher sensitivity, specificity and can detect COVID-19 with 91.7% accuracy23. Reference24 shows a deep learning model which obtains 92.4% results in detection of COVID-19. A ResNet50 model was proposed in25 which also achieved 98% results as well. All of these trials, nevertheless, took more time to diagnose and didn't produce the best outcomes because of information loss during the acquisition process. There are many studies on detection of COVID-19 that employ machine learning models with CT images26,27,28,29.A study presented in30proposes two different approaches with two systems each to diagnose tuberculosis from two datasets. In this study,initially, PCA) algorithm was employedto reduce the features dimensionality, aiming to extract the deep features. Then, SVM algorithm was used to for classifying features. This hybrid approachachieved an accuracy of 99.2%, a sensitivity of 99.23%, a specificity of 99.41%, and an AUC of 99.78%. Similarly, a study presented in31 utilizes different noise reduction techniques and compared the resultsby calculating qualitative visual inspection and quantitative parameters like Peak Signal-to-Noise Ratio (PSNR), Correlation Coefficient (Cr), and system complexity to determine the optimum denoising algorithm to be applied universally. However, these techniques manipulate all pixels whether they are contaminated by noise or not.An automated deep learning approach from Computed Tomography (CT) scan images to detect COVID-19 is proposed in32. In this method anisotropic diffusion techniques are used to de-noised the image and then CNN model is employed to train the dataset. At the end, different models including AlexNet, ResNet50, VGG16 and VGG19 have been evaluated in the experiments. This method worked well and achieved higher accuracy. However, when the images were contaminated with higher noise density, its performance suffered.Similarly, the authors in33 used four powerful pre-trained CNN models, VGG16, DenseNet121, ResNet50,and ResNet152, for the COVID-19 CT-scan binary classification task. In this method, a FastAI ResNet framework was designed to automatically find the best architecture using CT images. Additionally, a transfer learning techniques were used to overcome the large training time. This method achieved a higher F1 score of 96%. A deep learning method to detect COVID-19 using chest X-ray images was presented in 34. A dataset of 10,040 samples were used in this study. This model has a detection accuracy of 96.43% and a sensitivity of 93.68%.However, its performance dramatically decreases with higher density Poisson noise. A convolution neural networks method used for binary classification pneumonia-based conversion of VGG-19, Inception_V2, and decision tree model was presented in35. In this study, X-ray and CT scan images dataset that contains 360 images were used for COVID-19 detection. According to the findings, VGG-19, Inception_V2 and decision tree model illustrate high performance with accuracy of 91% than Inception_V2 (78%) and decision tree (60%) models.

In this paper, a paradigm for automatic COVID-19 screening that is based on assessment fusion is proposed. The effectiveness and efficiency of all baseline models were improved by our proposed model, which utilized the majority voting prediction technique to eliminate the mistakes of individual models. The proposed AFM model only needs chest X-ray images to diagnose COVID-19 in an accurate and speeding way.

The rest of the paper is organized as: The dataset is explained in section "Meterial and methods". section "Proposed method" explains our proposed approach and section "Results and Discussion" presents empirical results and analysis. section "Conclusion" describes conclusion and the specific contributions along with the future directions for improving the efficiency of the proposed work.

The rest is here:
A novel CT image de-noising and fusion based deep learning ... - Nature.com