[Paper Review] Lesion detection and Grading of Diabetic Retinopathy via Two-stages Deep Convolutional Neural Networks

Yehui Yang, Tao Li+4 more|arXiv (Cornell University)|May 1, 2017

Retinal Imaging and Analysis22 citations

TL;DR

This paper proposes a two-stage deep convolutional neural network (DCNN) for automated diabetic retinopathy (DR) analysis, combining lesion detection and severity grading. It uses a local network for lesion localization and a global network with an imbalanced weighting map to prioritize lesion-rich regions, significantly improving DR grading accuracy over end-to-end models.

ABSTRACT

We propose an automatic diabetic retinopathy (DR) analysis algorithm based on two-stages deep convolutional neural networks (DCNN). Compared to existing DCNN-based DR detection methods, the proposed algorithm have the following advantages: (1) Our method can point out the location and type of lesions in the fundus images, as well as giving the severity grades of DR. Moreover, since retina lesions and DR severity appear with different scales in fundus images, the integration of both local and global networks learn more complete and specific features for DR analysis. (2) By introducing imbalanced weighting map, more attentions will be given to lesion patches for DR grading, which significantly improve the performance of the proposed algorithm. In this study, we label 12,206 lesion patches and re-annotate the DR grades of 23,595 fundus images from Kaggle competition dataset. Under the guidance of clinical ophthalmologists, the experimental results show that our local lesion detection net achieve comparable performance with trained human observers, and the proposed imbalanced weighted scheme also be proved to significantly improve the capability of our DCNN-based DR grading algorithm.

Motivation & Objective

• To develop an automated system that detects retinal lesions and grades DR severity in fundus images, addressing limitations of existing end-to-end deep learning models.
• To overcome the challenge of small, inconspicuous lesions (e.g., microaneurysms) being overlooked in end-to-end DCNNs by introducing a two-stage architecture.
• To improve DR grading performance by applying an imbalanced weighting map that emphasizes lesion-dense regions during training.
• To provide clinically interpretable outputs by localizing lesion types (e.g., microaneurysms, hemorrhages) and assigning severity grades (0–3 for NPDR).

Proposed method

• The method employs a two-stage DCNN: a local network for lesion detection and a global network for DR severity grading.
• Lesion patches are extracted and labeled from 12,206 patches, and DR grades are re-annotated on 23,595 fundus images with clinical ophthalmologist guidance.
• An imbalanced weighting map is introduced to assign higher attention to image regions with more severe lesions, enhancing feature learning in the global grading network.
• The local network achieves performance comparable to trained human observers in binary lesion detection (AUC = 0.9590 for high-sensitivity/high-specificity points).
• The global network is trained using weighted loss to prioritize misclassified or subtle cases, improving robustness in early-stage DR detection.
• The framework is evaluated on a re-annotated dataset from the Kaggle Diabetic Retinopathy Detection competition, with validation via ROC curves and Kappa scores.

Experimental results

Research questions

• RQ1 Can a two-stage DCNN architecture outperform end-to-end DCNNs in detecting retinal lesions and grading DR severity?
• RQ2 Does the use of an imbalanced weighting map significantly improve the performance of DR grading networks compared to standard training?
• RQ3 Can the proposed method localize and classify multiple lesion types (e.g., microaneurysms, hemorrhages, exudates) while assigning accurate severity grades?
• RQ4 How does the performance of the lesion detection network compare to that of trained human observers on the same dataset?
• RQ5 To what extent does the weighting scheme reduce misclassification of mild or early-stage DR cases?

Key findings

• The proposed two-stage DCNN achieved an AUC of 0.9590 in referable DR detection using the weighted scheme, compared to 0.7986 with the non-weighted baseline.
• The lesion detection network achieved performance comparable to trained human observers, with sensitivity and specificity points marked on the ROC curve.
• The imbalanced weighting map significantly improved both Kappa score and accuracy in DR grading, as demonstrated in Fig. 7.
• The global network with the weighting scheme outperformed both the non-weighted version and the standard AlexNet in severity grading tasks.
• The integration of local and global networks reduced errors from the local network by leveraging holistic image context in the global grading step.
• The method successfully localized and classified three major lesion types—microaneurysms, hemorrhages, and exudates—while assigning accurate NPDR severity grades (0–3).