[Paper Review] Classification of Diabetic Retinopathy Severity in Fundus Images with DenseNet121 and ResNet50

Jonathan Zhang, Bowen Xie+3 more|arXiv (Cornell University)|Aug 19, 2021

Retinal Imaging and Analysis2 references4 citations

TL;DR

This study evaluates DenseNet121 and ResNet50 for classifying diabetic retinopathy severity in fundus images using RGB, green-filtered, and high-contrast image variants. It achieves a maximum validation accuracy of 85% and lowest validation loss of 0.17 with DenseNet121 on RGB images, demonstrating that model architecture and image preprocessing significantly influence performance in DR classification.

ABSTRACT

In this work, deep learning algorithms are used to classify fundus images in terms of diabetic retinopathy severity. Six different combinations of two model architectures, the Dense Convolutional Network-121 and the Residual Neural Network-50 and three image types, RGB, Green, and High Contrast, were tested to find the highest performing combination. We achieved an average validation loss of 0.17 and a max validation accuracy of 85 percent. By testing out multiple combinations, certain combinations of parameters performed better than others, though minimal variance was found overall. Green filtration was shown to perform the poorest, while amplified contrast appeared to have a negligible effect in comparison to RGB analysis. ResNet50 proved to be less of a robust model as opposed to DenseNet121.

Motivation & Objective

• To develop a deep learning model for automated classification of diabetic retinopathy (DR) severity in fundus images to address the shortage of human assessors.
• To evaluate the performance of two deep learning architectures—DenseNet121 and ResNet50—on different image preprocessing variants.
• To determine the optimal combination of model architecture and image type (RGB, green-filtered, high-contrast) for DR severity classification.
• To assess the impact of image preprocessing techniques such as green filtering and histogram equalization on model performance.
• To provide a foundation for scalable, accurate DR screening tools using transfer learning and convolutional neural networks.

Proposed method

• Preprocessed 3,662 fundus images from the APTOS Kaggle dataset by resizing to 244×244 pixels and applying three image types: original RGB, green-only filter (zeroing red and blue channels), and high-contrast (histogram-equalized RGB channels).
• Trained two deep learning models—DenseNet121 and ResNet50—using transfer learning with ImageNet weights, excluding the top classification layer.
• Constructed a sequential model with a global average pooling layer, dropout (0.5), and a 5-way softmax output layer for multi-class DR severity classification (0–4).
• Optimized models using Adam optimizer with a low initial learning rate (0.0001), binary cross-entropy loss, and accuracy as the metric.
• Applied data augmentation via Keras ImageDataGenerator with random zoom, horizontal and vertical flips, and constant padding to improve generalization.
• Split data into 80% training and 20% testing, and evaluated performance using validation loss and accuracy over 15 training epochs.

Experimental results

Research questions

• RQ1 Which combination of model architecture (DenseNet121 vs. ResNet50) and image preprocessing (RGB, green, high-contrast) yields the highest validation accuracy for DR severity classification?
• RQ2 How does green filtering of fundus images affect the performance of deep learning models in detecting diabetic retinopathy?
• RQ3 Does enhancing image contrast via histogram equalization significantly improve classification performance compared to standard RGB images?
• RQ4 How do the validation loss and accuracy of DenseNet121 and ResNet50 compare across different image preprocessing strategies?
• RQ5 To what extent does model architecture influence robustness and generalization in DR classification with limited training data?

Key findings

• The combination of DenseNet121 with RGB images achieved the lowest average validation loss of 0.17, indicating superior optimization stability.
• ResNet50 + RGB achieved the highest maximum validation accuracy of 85%, outperforming other combinations in classification performance.
• Green-filtered images yielded the poorest performance across all models, suggesting that color channel reduction diminishes critical diagnostic features.
• High-contrast images showed negligible improvement over standard RGB images, indicating that histogram equalization did not significantly enhance model performance.
• DenseNet121 demonstrated greater robustness than ResNet50, as it achieved lower loss and more consistent performance across image types.
• Despite high accuracy in most configurations, the overall variance in performance was minimal, suggesting that model choice and preprocessing have limited impact under current data constraints.