RetinoNet: an EfficientNet and feature pyramid network-based framework for accurate diabetic retinopathy classification

Background Damage to the blood vessels in the retina of the eye causes diabetic retinopathy (DR), one of the most common complications of diabetes. It is a major cause of vision impairment and blindness in several individuals. Due to the large number of cases, there is a lack of ophthalmologists performing conventional procedures, leading to a big challenge. Therefore, an efficient and scalable DR detection method is required. Methods This study introduced a hybrid model called RetinoNet, which is used to detect the severity of DR by minimizing the need for complex diagnostic procedures by ophthalmologists. First, the model uses a long chain of pre-processing steps consisting of adaptive histogram equalization, nonlocal means denoising, GrabCut for background masking, and simulated super-resolution to improve the resolution, particularly to extract small details in a given image, thereby partially addressing generalization when working with limited datasets. It then uses EfficientNet-B0 as the backbone for high-level feature extraction, along with a feature pyramid network (FPN) for multi-scale feature extraction, which helps increase the detection accuracy when an object of different scales is present in the image. Global average pooling (GAP) is used for dimension reduction and to prevent overfitting by minimizing the parameters for classification. Results RetinoNet captured fine details that were used to enhance the detection of small and large lesions, thereby achieving an accuracy of 96.8% on the Messidor dataset. It has been validated on the APTOS dataset with 2,400 imbalanced data samples and achieved 91.6% accuracy. To evaluate the results of pre-processing and classification, the model uses the image measurement metrics such in peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean squared error (MSE) with metrics such as the F1-score, precision, recall, and specificity. Across all classes, an average receiver operating characteristic (ROC) of 97.5% was obtained; the confusion metric findings showed fewer false positives and false negatives. Moreover, the analysis of variance (ANOVA) test on the picture measurements revealed bias-free image enhancement.