[Paper Review] Deep Learning based Early Detection and Grading of Diabetic Retinopathy Using Retinal Fundus Images

Sheikh Muhammad Saiful Islam, Md Mahedi Hasan+1 more|arXiv (Cornell University)|Dec 27, 2018

Retinal Imaging and Analysis72 citations

TL;DR

A CNN-based model for early-stage detection and severity grading of diabetic retinopathy on retinal fundus images, achieving state-of-the-art grading performance and high early-detection sensitivity. Evaluated on Kaggle EyePACS with 0.851 quadratic weighted kappa and 0.844 AUC; 98% sensitivity for early-stage detection.

ABSTRACT

Diabetic Retinopathy (DR) is a constantly deteriorating disease, being one of the leading causes of vision impairment and blindness. Subtle distinction among different grades and existence of many significant small features make the task of recognition very challenging. In addition, the present approach of retinopathy detection is a very laborious and time-intensive task, which heavily relies on the skill of a physician. Automated detection of diabetic retinopathy is essential to tackle these problems. Early-stage detection of diabetic retinopathy is also very important for diagnosis, which can prevent blindness with proper treatment. In this paper, we developed a novel deep convolutional neural network, which performs the early-stage detection by identifying all microaneurysms (MAs), the first signs of DR, along with correctly assigning labels to retinal fundus images which are graded into five categories. We have tested our network on the largest publicly available Kaggle diabetic retinopathy dataset, and achieved 0.851 quadratic weighted kappa score and 0.844 AUC score, which achieves the state-of-the-art performance on severity grading. In the early-stage detection, we have achieved a sensitivity of 98% and specificity of above 94%, which demonstrates the effectiveness of our proposed method. Our proposed architecture is at the same time very simple and efficient with respect to computational time and space are concerned.

Motivation & Objective

• Develop an automated deep learning model to detect early-stage diabetic retinopathy from retinal fundus images by identifying microaneurysms and related features.
• Grade DR severity into five categories using a regression-based output and mean squared error objective.
• Enhance performance through data preprocessing and extensive augmentation to handle imbalanced classes and image quality variations.

Proposed method

• Proposed an 18-layer deep CNN with 4x4 kernels and 3x3 max-pooling, trained with SGD and Nesterov momentum.
• Images preprocessed to standardize lighting and color per Graham's guidelines, then augmented with rotations, flips, translations, zooms, crops, and Krizhevsky color augmentation.
• Two fully connected layers (1024 units each) followed by a single output neuron; regression framing with mean squared error; loss clipped to range 0-4.
• Feature blending network combines features from multiple augmentations and both eyes using a small fully connected/maxout architecture.
• Training used orthogonal initialization, L2 regularization, dropout, and staged learning rates; blending network trained with Adam.
• Evaluation on EyePACS (Kaggle) with grade distribution and thresholds for discretizing regression outputs (0.5, 1.5, 2.5, 3.5).

Experimental results

Research questions

• RQ1 Can a CNN detect early-stage DR by identifying microaneurysms and related signs with high sensitivity and specificity?
• RQ2 How well can a regression-based CNN framework grade DR severity across five categories, and how does it compare with standard metrics like kappa and AUC?
• RQ3 What is the impact of data preprocessing and augmentation on model generalization for imbalanced DR grade distributions?
• RQ4 Does blending features from multiple augmentations and eyes improve severity grading performance?
• RQ5 What are the optimal training hyperparameters and architecture choices for robust DR detection and grading on the EyePACS dataset?

Key findings

• Achieved 98% sensitivity and >94% specificity for early-stage detection.
• Obtained 0.851 quadratic weighted kappa for severity grading on the Kaggle EyePACS test set.
• Achieved 0.844 AUROC and 0.743 F-score for grading performance.
• Used a simple and efficient 18-layer CNN with 4x4 kernels and regression-based output for grades 0–4.
• Data augmentation and preprocessing were crucial to prevent overfitting and handle imbalanced classes.
• Blending features from multiple augmentations and both eyes provided performance gains.