Deep Learning Approaches for Retinal Image Classification: A Comparative Study of GoogLeNet and ResNet Architectures

Abstract

This study evaluates the performance of four deep learning models, namely GoogLeNet (InceptionV3), ResNet-18, ResNet-50, and ResNet-101, in classifying Optical Coherence Tomography (OCT) images. Images were pre-processed by resizing them to 224x224 pixels and normalizing the pixel values. The models were fine-tuned using pre-trained weights from ImageNet dataset and trained for 10 iterations using categorical_crossentropy loss function and Adam optimizer. Performance metrics such as accuracy, precision, recall, specificity, and F1 score were calculated for each model. The results show that ResNet-101 outperforms other models with 96.69% accuracy, 96.85% sensitivity, and 98.90% specificity. ResNet-50 also showed high performance, while ResNet-18 showed the lowest performance with 33.99% accuracy. GoogLeNet achieved moderate results with 72.21% accuracy. ROC curves and confusion matrices are used to visualize the classification performance. ResNet-101 and ResNet-50 show superior performance in all classes, while ResNet-18 and GoogLeNet have higher misclassification rates. This study highlights the importance of model depth and residual connections in improving the classification performance of OCT images. The findings show that deeper models such as ResNet-50 and ResNet-101 are more effective in capturing complex features, leading to better classification accuracy.

This study evaluates the performance of four deep learning models, namely GoogLeNet (InceptionV3), ResNet-18, ResNet-50, and ResNet-101, in classifying Optical Coherence Tomography (OCT) images. Images were pre-processed by resizing them to 224x224 pixels and normalizing the pixel values. The models were fine-tuned using pre-trained weights from ImageNet dataset and trained for 10 iterations using categorical_crossentropy loss function and Adam optimizer. Performance metrics such as accuracy, precision, recall, specificity, and F1 score were calculated for each model. The results show that ResNet-101 outperforms other models with 96.69% accuracy, 96.85% sensitivity, and 98.90% specificity. ResNet-50 also showed high performance, while ResNet-18 showed the lowest performance with 33.99% accuracy. GoogLeNet achieved moderate results with 72.21% accuracy. ROC curves and confusion matrices are used to visualize the classification performance. ResNet-101 and ResNet-50 show superior performance in all classes, while ResNet-18 and GoogLeNet have higher misclassification rates. This study highlights the importance of model depth and residual connections in improving the classification performance of OCT images. The findings show that deeper models such as ResNet-50 and ResNet-101 are more effective in capturing complex features, leading to better classification accuracy.