Optimizing Diabetic Retinopathy Classification with Transfer Learning: A Lightweight Approach Using Model Clustering

Abstract

Accurate and rapid classification of diabetic retinopathy is critically significant in order to prevent vision loss. The advent of artificial intelligence introduces novel and potent methodologies for enhancing the classification of diabetic retinopathy as derived from medical imaging. Due to the large size of the model make it unsuitable in real world. This research paper is dedicated to the classification of diabetic retinopathy utilizing constrained resources while achieving elevated accuracy levels. We implemented weighted clustering technique within deep convolutional neural networks and transfer learning architectures: VGG 19, DenseNet 121, and EfficientNet B6. To mitigate the challenge posed by considerable model sizes without sacrificing accuracy, the best fit results were observed with EfficientNet-B6, where applying weighted clustering reduced the model size by a factor of 12 while maintaining high accuracy results of 92% for the APTOS-2019 data. This underscores the efficacy of employing lightweight techniques to enhance the practicality of extensive models for the early diagnosis of diabetic retinopathy.

Keywords

• diabetic retinopathy
• deep convolutional neural network
• transfer learning
• weighted clustering