Denoising of COVID-19 CT and chest X-ray images using deep learning techniques for various noises using single image

Abstract

The onset of COVID-19 pandemic had a great impact on the health, economy and livelihood of so many lives around the world. Early identification of positive cases and isolation followed by treatment is very crucial in order to receive prompt treatment and prevent the virus from spreading further. Chest X-ray (CXR) and Computed Tomography (CT) are widespread and cost-effective medical imaging radiographic tools which are presently used for diagnosing the covid-19 quickly. A crucial step towards establishing a quick COVID-19 pre-diagnosis and reducing the workload on medical professionals is the use of deep learning algorithms to identify positive CXR and CT pictures of infected individuals. However, the CXR images have complicated edge structures and rich texture details that are sensitive to noise, which can interfere with the machines and doctors diagnosis. This paper presents two state-of-art denoising techniques, Noise2Noise(N2N) and Noise2Void(N2V),to eliminate the noises that were added to COVID-19 chest x-ray scan image and CT medical image modalities by additive Gaussian noise, Speckle noise, and salt-and-pepper noise. These two techniques do not require a pair of noisy and clean photos; instead, they denoise a single noisy image. Based on the study, Noise2Void performs well in removing Gaussian, Speckle, and salt-and-pepper noise from CXR image modalities. Similarly, Noise2Noise performance is good to remove only Gaussian noise in CT images. In the case of Speckle and salt and pepper noise in CT images. Noise2void gives better quality images with better PSNR and SSIM. The results are measured quantitatively and qualitatively using Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure(SSIM). In this paper, we experiment two learning based methods to de-noise images with high noise. The proposed method is beneficial for applications involving images that are highly prone to noise.