A Hybrid-Based Generative Adversarial Network Pattern For Diabetic Retinopathy Prediction

Abstract

Diabetic retinopathy (DR), a primary cause of lack of vision, can be detected early with comprehensive examination of fundus visuals of the retina that can be averted worldwide. Several noise sources frequently damage fundus visuals, reducing visual quality and making it more difficult to see fragile macular structures like tiny blood vessels and microaneurysms. On tiny medical information sets, noise reduction methods powered by deep training are computationally challenging and prone to excessive fitting. To address these issues, the current study suggests a hybrid–classical construction called the QGAN, which employs hybrid circuits in latent space in conjunction with convolutional component coding. By enhancing the latent representations through hybrid superposition and entanglement, the recommended QGAN enhances noise reduction and macular detail maintenance. In tests on the DR-based Kaggle information set, QGAN outperforms better than conventional noise reduction systems like CAE, ResNet, and DnCNN thanks to a PSNR of 39 dB, SSIM of 0.97, and AMI of 0.89. Although there is a little computing cost related to shallow hybrid circuits, the method maintains sensitive macular patterns and consistency in intensity. The aforementioned findings show that QGAN is a viable pretreatment method in early diabetic retinopathy and a promising hybrid-aided construction for noise reduction macular visuals.