Optimized CNN for Multi-Class Retinal Disease Detection with Minimized Memory Usage

Aim: The research aims to develop a Convolutional Neural Network optimized for multiclass retinal illness detection with high diagnostic accuracy while using the least amount of memory possible. Materials and Methods: An open-access collection of fundus photos used in the study has annotations for AMD, glaucoma, diabetic retinopathy and normal. Group 1: Conventional deep CNN designs such as VGG16, ResNet, Inception, and DenseNet are among the existing techniques for multiclass retinal disease diagnosis using CNN. The sample size is a total of 10,000 retinal images. Group 2: To minimize memory consumption while preserving high accuracy, the CNN architecture is optimized by the use of depth-wise separable convolutions, pruning, and quantization. Result: An accuracy of 92% was accomplished at the reduced memory usage of 45%, and fast inference per image was performed with an average duration of 200 ms. Accuracy in terms of precision for different disease diagnoses could reach 0.90 for diabetic retinopathy, 0.89 for AMD, and 0.93 for glaucoma with statistically significant performance where the achieved p-value is below 0.005. Conclusion: Using memory-efficient approaches, the study successfully shows that CNN can efficiently categorize various retinal disorders while using the least amount of memory possible.