A novel CT image de-noising and fusion based deep learning network to screen for disease (COVID-19)

Abstract

A COVID-19, caused by SARS-CoV-2, has been declared a global pandemic by WHO. It first appeared in China at the end of 2019 and quickly spread throughout the world. During the third layer, it became more critical. COVID-19 spread is extremely difficult to control, and a huge number of suspected cases must be screened for a cure as soon as possible. COVID-19 laboratory testing takes time and can result in significant false negatives. To combat COVID-19, reliable, accurate and fast methods are urgently needed. The commonly used Reverse Transcription Polymerase Chain Reaction has a low sensitivity of approximately 60% to 70%, and sometimes even produces negative results. Computer Tomography (CT) has been observed to be a subtle approach to detecting COVID-19, and it may be the best screening method. The scanned image's quality, which is impacted by motion-induced Poisson or Impulse noise, is vital. In order to improve the quality of the acquired image for post segmentation, a novel Impulse and Poisson noise reduction method employing boundary division max/min intensities elimination along with an adaptive window size mechanism is proposed. In the second phase, a number of CNN techniques are explored for detecting COVID-19 from CT images and an Assessment Fusion Based model is proposed to predict the result. The AFM combines the results for cutting-edge CNN architectures and generates a final prediction based on choices. The empirical results demonstrate that our proposed method performs extensively and is extremely useful in actual diagnostic situations.

Figure 1

Corona virus structure.

Figure 2

Shows the normal and noisy CT images. (a) normal CT image having natural Impulse noise (b) Impulse noise corrupted image with 50% noise density (c) Impulse noise corrupted image with 80% noise (d) Poisson noise corrupted image with 50% noise density (e) Poisson noise corrupted image with 80% noise density (f) contaminated image with both Impulse and Poisson noise with 80% noise density.

Figure 3

Samples of COVID19 enhanced CT images.

Figure 4

Representation of MSE and PSNR values for various windows.

Figure 5

Proposed Assessment Based Fusion Model.

Figure 6

Empirically Assessment of Fusion Based Model.

Figure 7

Structure Similarity Index (SSIM) values of all evaluated approaches. Noise density is from 55 to 90%. BDND = boundary discriminative noise detection, SR = square-root.

Figure 8

Mean Square Error (MSE) values of all evaluated approaches. Noise density is from 5 to 90%. BDND = boundary discriminative noise detection, SR = square-root.

Figure 9

shows the normal, noisy and de-noised CT images. (a) Normal CT image having natural Impulse noise (b) Impulse noise corrupted image with 80% noise density (c) De-noised image with SM (d) De-noised image with BDND (e) De-noised image with proposed method.

Figure 10

Average behavior of each and every model with 90% of confidence intervals.

Figure 11

Sensitivity and specificity of all models.

Figure 12

Average Precession and recall of all models.