Quality Assessment, Enhancement and Active Retraining for Reliable Fundus Image Analysis

Abstract

Fundus photography is widely used for retinal screening, but poor image quality caused by uneven lighting, blur, or acquisition errors often makes a large number of images clinically unusable. To address this, we propose a combined framework that first assesses image quality using deep learning known as Image Quality Assessment (IQA) model and then applies targeted enhancement to recover low-quality images. We trained and compared three convolutional neural networks MCFNet, ResNet50, and EfficientNet-B3 to classify fundus images as Good, Usable, and Bad using the EyeQ dataset. MCFNet performed best, reaching 93.78% test accuracy, largely due to its multi-color space fusion design. Rather than simply discarding poor-quality images, we applied CLAHE and Unsharp Masking to enhance marginal images and then retrained the IQA model on this improved data. This approach recovered around 45% of previously rejected images. On an external validation set, the retrained model reclassified 84 low-quality images into the Good category. Our results indicate that combining quality assessment with active enhancement can meaningfully reduce patient recall rates in tele ophthalmology, making screening workflows more practical and efficient.