Enhanced microaneurysm detection in fundus images using advanced image processing and machine learning techniques

Abstract

Diabetic retinopathy (DR) is a major cause of vision loss globally, with microaneurysms (MAs) often indicating early disease onset. Traditional MA detection involves manual examination of fundus images by ophthalmologists, which is labor-intensive, subjective, and variable. This study explores advanced image processing and machine learning techniques to develop an automated and objective method for improved MA detection in fundus images. The methodology involves key steps: image preprocessing using contrast-limited adaptive histogram equalization (CLAHE) and gamma correction to enhance MA visibility; 2D convolution filtering with an 11x11 kernel to extract MA features. Otsu's global thresholding to binarize the image; and morphological operations like top-hat and opening to remove noise and refine the binary mask. Potential MA candidates are identified, and a convolutional neural network (CNN) inspired by U-Net classifies them as true MAs or non-MAs. Trained on fundus images with expert annotations, the CNN achieved 92% accuracy, 88% sensitivity, and 94% specificity on the DRD-EPCC dataset. Future research should address image quality variability and explore advanced denoising and contrast methods for a comprehensive DR screening solution. This work highlights the potential of leveraging image processing and machine learning for automated, accurate MA detection, aiding earlier DR diagnosis and treatment