Adaptive noise and attention mechanisms based light-weight U-net based architecture for robust medical image denoising of retinal fundus images

Diabetic Retinopathy (DR) is a complication of diabetes, one of the leading causes of vision impairment and eventual blindness. This crucial eye condition called DR affects the light-sensitive tissues present at the back of the eye, causing damage to nerves and vision loss. Separating between the normal lesions and nerves from the affected ones remains a challenge, causing missed chances in diagnosing and treating DR. This is due to the noise present within the retinal images used for DR diagnosis. As a step toward supporting downstream DR analysis, this work addresses the underlying image-quality problem by proposing a Deep Learning-based image denoising model called the Noise-Aware Residual U-Net (NAR-UNet) architecture. The STARE dataset is employed in this study, and five-fold cross-validation is applied with 80% training and 20% validation data. The Retinal Fundus Images (RFI) are subjected to preprocessing followed by Noise Injection using the Adaptive Noise Module from the proposed framework, and subsequently the model is trained on the Noisy and clean RFI pairs considering the hybrid loss for learning, which is a combination of Mean Squared Error(MSE), Structural Similarity Index Measure (SSIM) and Sobel-based Gradient Loss. The experimental results have shown the best performance with a five-fold cross-validation mean Peak Signal-to-Noise Ratio (PSNR) of 37.99 dB, an SSIM of 0.946, and an average pixel accuracy of 0.9963 when compared with other state-of-the-art methods. The proposed model was also experimented on the Messidor dataset for the evaluation of the model’s generalizability and obtained a satisfactory five-fold cross-validation mean result with a PSNR of 40.46 dB, a SSIM of 0.953, and an average pixel accuracy of 0.9990. Thus, this paper demonstrates a deep learning framework for effective denoising of noisy retinal images; the resulting image-quality metrics (PSNR, SSIM, pixel accuracy, EME) characterize denoising performance only, and the model’s effect on downstream DR diagnosis, lesion detection, or vessel segmentation has not been experimentally evaluated in this study.