Medical image Denoising using Scaled Dual Convolutional Neural network (SD-CNN)

In medical image analysis, image denoising is a critical pre-processing step. Various algorithms with various denoising performances have been proposed during the last three decades. Deep learning-based models have recently showed enormous potential, outperforming all conventional methods. However, these methods are hampered by the need for a large training sample size and high computing costs. We show in this research a novel method built using convolutional layers which can be employed for efficient denoising of medical images with limited sample sizes. We provide a novel method for achieving the following goals by combining the qualities of various algorithms in this research. (1) Improved performance (2) Faster processing (3) Reduced denoiser complexity (4) Extracting complementary features to increase the denoiser's total capacity. We tested our method on retinal images (STARE dataset) in terms of qualitative and quantitative analysis. According on simulation results, the proposed technique Scaled Dual CNN (SD-CNN) outperforms earlier methods.

I. Introduction

One of the rising scientific disciplines is biomedical image analysis. It is focused with the study and analysis of digital images using image processing techniques and computational tools to aid in the diagnosis of clinical problems. Rapid advancements in biomedical image processing have lately been demonstrated to be critical, as they minimize the need for invasive diagnostic procedures. Computer Aided Diagnosis (CAD) is a popular tool for detecting and diagnosing disorders such as tumors, malignancies, and edemas. The examination of fundus images aids in the diagnosis of severe retinal illnesses such as diabetic retinopathy, macular degeneration, retinitis pigmentosa, and others. Images taken with a standard fundus camera do not create high-quality retinal images. The photos are primarily hampered by noise, low contrast features, and uneven illumination. To detect and analyze tiny changes in retinal vasculature and to use disease detection algorithms, the fundus image must be efficiently denoised. High blood pressure and diabetes are the leading causes of well-known eye disorders such as glaucoma and diabetic retinopathy. These disorders can progress to advanced stages without causing significant symptoms, but common symptoms include intraretinal microvascular anomalies and leakages. Fundus photography is critical in the detection and treatment of ophthalmologic diseases such as Age-Related Macular Degeneration (AMD) and Diabetic Retinopathy. Imperfections in the fundus camera optics, human eye aberrations, improper camera calibration, and flash lighting are all significant causes of retinal image quality decline. The main problem with noise is that it is inherently random and hence it cannot be completely removed from an image. The CCD, not the system electronics, limits the noise performance of an ideal digital camera. There are two types of noise sources: temporal noise and spatial noise. Temporal noise includes shot noise, reset noise, output amplifier noise, and dark current shot noise. Spatial noise include distortion, which is stable in time over display area.