[Paper Review] Conversion and Implementation of State-of-the-Art Deep Learning Algorithms for the Classification of Diabetic Retinopathy

M V Achuth Rao, Michelle Zhu+1 more|arXiv (Cornell University)|Oct 7, 2020

Retinal Imaging and Analysis13 references9 citations

TL;DR

This study evaluates state-of-the-art convolutional neural networks (CNNs) for diabetic retinopathy (DR) classification using transfer learning and data augmentation on limited, imbalanced medical retinal images. ResNet50 with Adam optimizer achieved the highest performance in binary classification (85.02% accuracy), while InceptionResNetV2 with Adam excelled in multi-class classification (85.02% accuracy, 97% macro AUC).

ABSTRACT

Diabetic retinopathy (DR) is a retinal microvascular condition that emerges in diabetic patients. DR will continue to be a leading cause of blindness worldwide, with a predicted 191.0 million globally diagnosed patients in 2030. Microaneurysms, hemorrhages, exudates, and cotton wool spots are common signs of DR. However, they can be small and hard for human eyes to detect. Early detection of DR is crucial for effective clinical treatment. Existing methods to classify images require much time for feature extraction and selection, and are limited in their performance. Convolutional Neural Networks (CNNs), as an emerging deep learning (DL) method, have proven their potential in image classification tasks. In this paper, comprehensive experimental studies of implementing state-of-the-art CNNs for the detection and classification of DR are conducted in order to determine the top performing classifiers for the task. Five CNN classifiers, namely Inception-V3, VGG19, VGG16, ResNet50, and InceptionResNetV2, are evaluated through experiments. They categorize medical images into five different classes based on DR severity. Data augmentation and transfer learning techniques are applied since annotated medical images are limited and imbalanced. Experimental results indicate that the ResNet50 classifier has top performance for binary classification and that the InceptionResNetV2 classifier has top performance for multi-class DR classification.

Motivation & Objective

• To address the challenge of limited and imbalanced medical retinal image datasets for diabetic retinopathy (DR) classification.
• To evaluate state-of-the-art deep learning models—Inception-V3, VGG16, VGG19, ResNet50, and InceptionResNetV2—for DR image classification.
• To determine the optimal combination of model architecture and optimizer for both binary and multi-class DR severity classification.
• To explore the impact of input size and skip connections on model performance in medical image classification tasks.
• To propose a deployable, cost-effective deep learning diagnostic tool for DR detection, especially in resource-limited regions.

Proposed method

• Employed transfer learning using pre-trained ImageNet weights to initialize the CNN models, reducing training time and improving performance on small datasets.
• Applied data augmentation techniques such as rotation, scaling, and noise injection to increase training data diversity and mitigate class imbalance.
• Used two input sizes—224×224 and 299×299 pixels—evaluating model performance across different resolutions to assess input size impact.
• Trained models with two optimizers: Adam and Stochastic Gradient Descent (SGD), comparing their effects on convergence and accuracy.
• Implemented multi-class classification with five severity levels (0–4) based on DR progression, using categorical cross-entropy loss.
• Proposed a multi-stage binary classification system using a decision tree-like approach, where sigmoid outputs from individual binary models are combined via Argmax for final 5-class prediction.

Experimental results

Research questions

• RQ1 Which state-of-the-art CNN architecture achieves the highest classification accuracy for diabetic retinopathy in binary and multi-class settings?
• RQ2 How do different optimizers (Adam vs. SGD) affect model convergence and performance on limited medical imaging data?
• RQ3 Does using the model’s default input size improve performance compared to custom-sized inputs?
• RQ4 To what extent do skip connections in residual networks (e.g., ResNet50) mitigate the vanishing gradient problem and enhance learning?
• RQ5 Can a multi-stage binary classification system improve the accuracy of 5-class DR severity prediction compared to direct multi-class training?

Key findings

• ResNet50 with the Adam optimizer achieved the highest test accuracy of 85.02% in binary classification, outperforming other models.
• InceptionResNetV2 with Adam achieved the best performance in multi-class classification, reaching 85.02% test accuracy and a macro average AUC of 0.97.
• Models trained with Adam showed significantly better performance than those with Stochastic Gradient Descent, especially in multi-class tasks.
• The use of model-specific default input sizes (224×224 or 299×299) led to better performance than re-sized inputs, likely due to preserved pre-trained feature maps.
• Skip connections in ResNet50 and InceptionResNetV2 helped mitigate the vanishing gradient problem, enabling more stable and effective training on deep architectures.
• The confusion matrix and ROC curves confirmed high specificity and sensitivity for the best-performing models, particularly InceptionResNetV2 in multi-class settings.