[Paper Review] Deep Learning-Based Detection of Referable Diabetic Retinopathy and Macular Edema Using Ultra-Widefield Fundus Imaging

Philippe Zhang, Pierre-Henri Conze+3 more|arXiv (Cornell University)|Sep 19, 2024

Retinal Imaging and AnalysisMedicine3 citations

TL;DR

This paper proposes a deep learning framework using EfficientNet and ResNet architectures with multi-level feature extraction and test-time augmentation to automate detection of referable diabetic retinopathy and diabetic macular edema from ultra-widefield (UWF) fundus images. The best model achieved an AUROC of 0.9820 and AUPRC of 0.9699 on the MICCAI 2024 UWF4DR challenge test set, demonstrating high accuracy and robustness for clinical screening applications.

ABSTRACT

Diabetic retinopathy and diabetic macular edema are significant complications of diabetes that can lead to vision loss. Early detection through ultra-widefield fundus imaging enhances patient outcomes but presents challenges in image quality and analysis scale. This paper introduces deep learning solutions for automated UWF image analysis within the framework of the MICCAI 2024 UWF4DR challenge. We detail methods and results across three tasks: image quality assessment, detection of referable DR, and identification of DME. Employing advanced convolutional neural network architectures such as EfficientNet and ResNet, along with preprocessing and augmentation strategies, our models demonstrate robust performance in these tasks. Results indicate that deep learning can significantly aid in the automated analysis of UWF images, potentially improving the efficiency and accuracy of DR and DME detection in clinical settings.

Motivation & Objective

• To develop a robust deep learning system for automated analysis of ultra-widefield (UWF) fundus images to improve early detection of diabetic retinopathy (DR) and diabetic macular edema (DME).
• To address challenges in UWF image quality variability and the need for scalable, expert-independent screening in clinical settings.
• To enhance diagnostic accuracy and efficiency by leveraging transfer learning, multi-level feature extraction, and test-time augmentation.
• To evaluate performance across three tasks: image quality assessment, referable DR detection, and DME identification using standardized benchmarks.
• To support clinical deployment by balancing high sensitivity and specificity, minimizing false positives while maximizing detection of treatable cases.

Proposed method

• Employed EfficientNet-B0 and ResNet-50 architectures fine-tuned on UWF fundus images from the MICCAI 2024 UWF4DR challenge dataset.
• Applied a multi-level feature extraction strategy in the Multi-Level EfficientNet-B0 model to capture hierarchical patterns across different receptive field scales.
• Implemented a color normalization pipeline using Gaussian blurring and local mean subtraction to reduce color variability and enhance texture features.
• Used data augmentation techniques including random horizontal flipping, rotation, and mixup to improve model generalization and reduce overfitting.
• Applied test-time augmentation (TTA) by generating multiple augmented versions of each test image and averaging predictions to boost robustness.
• Trained models using the Adam optimizer with an initial learning rate of 1e-4 and an exponential learning rate scheduler, with transfer learning from pre-trained ImageNet weights.

Figure 1: Samples from the quality assessment dataset: (A) good quality, (B) bad quality.

Experimental results

Research questions

• RQ1 Can deep learning models effectively detect referable diabetic retinopathy in ultra-widefield fundus images with high sensitivity and specificity?
• RQ2 How does test-time augmentation improve the robustness and performance of deep learning models in DME detection from UWF images?
• RQ3 To what extent does multi-level feature extraction enhance the detection of subtle pathological signs in DR and DME compared to standard feature extraction?
• RQ4 Can transfer learning from pre-trained models significantly improve performance on limited medical imaging datasets for UWF fundus analysis?
• RQ5 How do preprocessing techniques like color normalization impact model generalization across diverse UWF image quality conditions?

Key findings

• The Multi-Level EfficientNet-B0 with test-time augmentation achieved the highest AUROC (0.9820) and AUPRC (0.9699) on the test set for DME detection, outperforming other models.
• The same model achieved a specificity of 0.9577, indicating strong performance in minimizing false positives while maintaining high detection sensitivity.
• The standard EfficientNet-B0 model achieved the highest sensitivity (0.9500), demonstrating strong ability to identify most true positive DME cases.
• The Multi-Level EfficientNet-B0 without TTA achieved a perfect specificity of 1.0000 on the validation set, highlighting its precision in excluding false positives.
• Fine-tuning from the referable DR detection task significantly improved performance on the DME detection task, indicating shared feature utility across related pathologies.
• The integration of multi-level feature extraction and TTA enhanced model robustness and generalization, suggesting strong potential for clinical deployment.

Figure 2: Preprocessing of images: (A) original image, (B) center-cropped image, (C) preprocessed image.