Noise and performance analysis on fundus images with CNN and transformer models

Fundus imaging is a valuable diagnostic tool in ophthalmology, providing clinicians with detailed visualizations of the retina and aiding in the detection and monitoring of various eye diseases, including age-related macular degeneration (AMD), glaucoma, diabetic retinopathy (DR), and cataract. However, the quality of fundus images can be significantly affected by noise, mainly additive white Gaussian noise (AWGN), which is inherent in many imaging systems. The presence of noise in real-world data poses significant challenges for computer vision tasks. In the field of medical image classification, a wrong diagnoisis has heavy consequences. Understanding the impact of AWGN on fundus images is crucial for developing practical denoising algorithms and improving diagnostic accuracy. This work presents an analysis of AWGN noise in fundus images aims to characterize its effects on image quality and assess its impact on diagnostic tasks. The work also analyzes the performance of six models (3 each) of two popular deep learning architectures, Convolutional Neural Networks (CNN) and Vision Transformers (ViT) in the presence of AWGN. AWGN is first introduced to the clean image datasets to conduct the analysis. The CNN and ViT models are trained on the noisy datasets to evaluate the performance of the image classification task. The work also involves six denoising algorithms and a popular image enhancement algorithm- Contrast Limited Adaptive Histogram Equalization (CLAHE).

I. Introduction

In practical scenarios, medical images are often affected by noise. Noise can distort critical features thereby affecting the quality of the perceived image and the accuracy of automated diagnosis and disease detection. Analysing noise and performance in fundus images is essential for analyzing the quality and reliability of diagnostic information derived from retinal fundus images. This paper explores the challenges, methods and significance of noise reduction and performance analysis in fundus image analysis. Fundus images are affected by both additive and multiplicative noise. The additive and multiplicative noises are often modelled in literature as AWGN and speckle noise respectively. While speckle noises reduce clarity to a greater extent, it is AWGN caused due to sensor limitations in low light conditions that is more prevalent. The paper explores the effect of AWGN on popular deep learning models, different methods like deep learning models, statistical methods and filtering methods to reduce AWGN present in fundus images as well as image enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE).