Retina Image Classification Using ML

Diabetic retinopathy (DR), an ocular complication of diabetes that can result in vision loss, requires prompt identification and precise assessment for effective prevention of blindness in diabetic patients. Contemporary approaches to DR detection predominantly utilize convolutional neural networks (CNNs) and deep learning architectures to differentiate blood vessel pixels within fundus images. This research introduces an enhanced activation function for DR diagnosis that automatically minimizes loss metrics while reducing computational time. Our methodology includes fine-tuning the model to enable lower-level layers to capture localized structures of both pathological and normal regions. Rather than employing traditional fully connected (FC) layers—which encode domain- specific global features—we implement a novel FC layer utilizing principal component analysis (PCA) in an unsupervised context to extract distinctive features from fundus imagery. This modification decreases model complexity and substantially mitigates overfitting issues while adapting the model to fundus image characteristics, thereby optimizing DR feature detection capabilities. The architecture culminates with a gradient-boosting classification layer. Performance evaluation through 10-fold cross-validation on two rigorous datasets (Eye PACS and Messidor) demonstrates superior results compared to existing methodologies. The system offers significant value for preliminary DR screening and assists medical professionals in expediting referral decisions for ophthalmological consultation, diagnosis, and treatment interventions. Keywords: Diabetic retinopathy, fundus images, CNNs, activation function.