An Effective Approach to Detect Lesions in Color Retinal Images

Huan Wang, Wynne Hsu, Kheng Guan Goh, Mong-Li Lee

Introduction

Diabetic retinopathy remains a primary cause of vision loss worldwide. Early intervention through regular screenings is effective but hampered by the massive volume of retinal fundus images that require expert manual review. Automating this process is challenging due to non-uniform illumination and the morphological variability of lesions (exudates). This paper presents a robust automated pipeline, the "Enhanced Minimum Distance Discriminant (MDD)," designed to filter healthy images and identify abnormalities with high sensitivity.

Methodology

The proposed Enhanced MDD pipeline consists of three critical stages to ensure reliable detection across varying image qualities.

1. Non-linear Brightness Adjustment

To mitigate uneven lighting—a common issue where the periphery of the retina appears darker than the center—the system applies a non-linear transformation: $y = \beta \cdot x^\alpha$. This adjustment selectively enhances darker regions without over-saturating well-lit areas, ensuring that peripheral lesions are not overlooked.

2. Statistical Classification (MDD)

The system converts images from the standard RGB color space into a spherical coordinate system $(L, \theta, \phi)$. Here, $L$ represents intensity, while $\theta$ and $\phi$ capture chromaticity. By employing a Bayesian-derived Minimum Distance Discriminant, the classifier compares pixel color vectors against established clusters for lesions and background tissues. This separation of intensity from color allows for more consistent detection across different lighting conditions.

3. Local-Window Verification

Brightness enhancement can occasionally amplify background noise, leading to false positives. To counter this, a verification step uses an $M \times M$ local window to calculate a contrast measure ($D$). Since true lesions typically possess sharper edges than the gradual gradients of the retinal background, this step filters out artifacts that lack the distinct boundaries of real exudates.

Results and Performance

The Enhanced MDD system was evaluated on a dataset of 154 color retinal images (54 abnormal, 100 normal) sourced from clinics in Singapore.

Detection of retinal lesions: the original fundus image (left) and the resulting binary mask identifying exudates (right).

The results demonstrated a significant improvement over baseline methods. While a standard MDD classifier missed approximately 33% of abnormal images, the Enhanced MDD achieved 100% sensitivity, identifying every image containing lesions. Furthermore, the system correctly classified 70% of normal images as healthy. In a clinical screening workflow, this translates to a 70% reduction in manual workload, as specialists only need to review the remaining 30% of "suspicious" healthy images and the confirmed abnormal cases.

Discussion and Future Work

A significant technical hurdle in retinal image analysis is the presence of the optic disc. This anatomical feature shares color characteristics with lesions, often leading to misclassification.

The challenge of the optic disc: Original image (top), mask of the optic disc (bottom-left), and the detected lesion mask (bottom-right).

While the current pipeline effectively isolates lesions, integrating specialized optic disc masking algorithms is a priority for future development. Additionally, expanding the dataset to include diverse ethnicities and camera types will be necessary to validate the system's generalizability. Future iterations may also move beyond binary classification to provide automated grading of retinopathy severity.

Conclusion

The Enhanced MDD pipeline offers a scalable solution for diabetic retinopathy screening. By combining intelligent brightness normalization with statistical color classification and local edge verification, the system achieves the 100% sensitivity required for medical applications while significantly reducing the burden on healthcare professionals.

Summary updated

Abstract

Diabetic-related eye diseases are the most common cause of blindness in the world. So far the most effective treatment for these eye diseases is early detection through regular screening. To lower the cost of such screenings, we employ state-of-the-art image processing techniques to automatically detect the presence of abnormalities in the retinal images obtained during the screenings. The authors focus on one of the abnormal signs: the presence of exudates/lesions in the retinal images. We propose a novel approach that combines brightness adjustment procedure with statistical classification method and local-window-based verification strategy. Experimental results indicate that we are able to achieve 100% accuracy in terms of identifying all the retinal images with exudates while maintaining a 70% accuracy in correctly classifying the truly normal retinal images as normal. This translates to a huge amount of savings in terms of the number of retinal images that need to be manually reviewed by the medical professionals each year.
BibTeX
@inproceedings{wang2000effective,
  author = {Huan Wang and Wynne Hsu and Kheng Guan Goh and Mong-Li Lee},
  title = {{An Effective Approach to Detect Lesions in Color Retinal Images}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2000},
  doi = {10.1109/cvpr.2000.854775},
}
Figure