Optimising Diabetic Retinopathy Classification using EfficientNetB7

Diabetic retinopathy (DR) is a significant cause of vision impairment and blindness in individuals with diabetes, necessitating early and accurate detection of its various stages for effective treatment. This study applies EfficientN etB7, an advanced deep learning model, to classify different stages of diabetic retinopathy and compares its performance with other established models such as DenseN et121, Inception V3,and VGG16. The study uses datasets like the Kaggle Diabetic Retinopathy Detection and Messidor datasets, which provide annotated retinal images across multiple stages of DR: mild, moderate, severe, proliferative diabetic retinopathy, and normal. Key preprocessing techniques are employed to enhance the quality of the input data. These include noise reduction, contrast enhancement using CLAHE, and data augmentation to enhance the model's reslience. EfficientNetB7 is selected due to its optimal network depth, width, and resolution balance, which facilitates high accuracy with fewer computational resources. The model's performance is evaluated using metrics including Fl score, accuracy, precision, and recall, which are crucial for assessing the reliability of medical image classification models. Comparative analysis reveals that EfficientN etB7 outperforms DenseNet121, InceptionV3, and VGG16 across all evaluation metrics. Specifically, EfficientN etB7 achieves superior accuracy, precision, recall, and Fl score in classifying the different stages of diabetic retinopathy. These results highlight the effectiveness of EfficientN etB7 in providing reliable and efficient detection of diabetic retinopathy, highlighting its potential for improving clinical outcomes through early and precise diagnosis.