Source-linked AI summary
An Ensemble-based System for Microaneurysm Detection and Diabetic Retinopathy Grading
Balint Antal, Andras Hajdu
TL;DR
Reliable microaneurysm detection remains difficult, yet it matters because microaneurysms are early diabetic-retinopathy signs and grading depends on their detection. The paper builds an ensemble by combining preprocessing methods with candidate extractors and evaluates it for detection and grading. The detector ranked first in an online competition, while Messidor grading achieved 0.90±0.01 AUC, although serious diabetic-retinopathy stages can still be misclassified.
Problem
Reliable microaneurysm detection is an open issue, and it matters because microaneurysms are early diabetic-retinopathy signs while grading depends on their detection.
Method
The framework combines preprocessing-method and candidate-extractor pairs, then selects and fuses an ensemble of these components.
Results
The detector ranked first in an online microaneurysm-detection competition and achieved 0.90±0.01 AUC on Messidor diabetic-retinopathy grading.
Takeaways & Limitations
The ensemble provides competitive microaneurysm detection and diabetic-retinopathy grading performance across the evaluated settings.
Takeaways & Limitations
The system still misclassifies some stages containing serious diabetic retinopathy, indicating that additional lesions such as exudates are needed for clinical grading.
Abstract
from arXiv · showhide
Reliable microaneurysm detection in digital fundus images is still an open issue in medical image processing. We propose an ensemble-based framework to improve microaneurysm detection. Unlike the well-known approach of considering the output of multiple classifiers, we propose a combination of internal components of microaneurysm detectors, namely preprocessing methods and candidate extractors. We have evaluated our approach for microaneurysm detection in an online competition, where this algorithm is currently ranked as first and also on two other databases. Since microaneurysm detection is decisive in diabetic retinopathy grading, we also tested the proposed method for this task on the publicly available Messidor database, where a promising AUC 0.90 with 0.01 uncertainty is achieved in a 'DR/non-DR'-type classification based on the presence or absence of the microaneurysms.
I. INTRODUCTION
Diabetic retinopathy screening needs reliable, scalable analysis because manual grading is slow and resource demanding. The paper targets microaneurysm detection by combining detector components rather than only classifier outputs.
- Manual grading is slow and resource demanding, motivating computer-aided screening from color fundus images.
- Microaneurysms are usually the earliest sign of diabetic retinopathy, making timely and precise detection essential.
- Computer-aided diabetic retinopathy grading depends strongly on microaneurysm detection.
- Standard ensemble methods combine class labels or real values, but microaneurysm detectors produce spatial candidate coordinates that may not coincide.
- The proposed framework combines preprocessing methods and candidate extractors, selects combinations, and evaluates both microaneurysm detection and diabetic retinopathy grading.
II. PREPROCESSING METHODS
The framework selects preprocessing methods that can precede different candidate extractors while preserving original image characteristics. It excludes methods that generate excessive noise for microaneurysm detection.
- Preprocessing methods are applied before microaneurysm candidate extraction.
- Selected methods must be interchangeable across candidate extractors and preserve the characteristics of the original images.
- Histogram equalization, adaptive histogram equalization, and color normalization were excluded because they generated overly noisy images for microaneurysm detection.
- The described contrast-enhancement method uses a gray-level transformation parameterized by the original image mean and a transition parameter r.
B. Contrast limited adaptive histogram equalization [10]
The preprocessing methods include local contrast enhancement, vessel removal with inpainting, and illumination equalization, each targeting a distinct image characteristic.
- B. Contrast limited adaptive histogram equalization [10]: CLAHE splits the image into disjoint regions and applies local histogram equalization in each region.
- B. Contrast limited adaptive histogram equalization [10]: Bilinear interpolation eliminates boundaries between the locally equalized regions.
- Removing the complete vessel system and filling its holes with inpainting makes microaneurysms near vessels more easily detectable.
- Illumination equalization reduces vignetting caused by uneven illumination and enhances microaneurysms appearing on the retinal border.
E. No preprocessing
The analysis includes candidate extraction on both preprocessed and original images. Candidate extractors identify image objects with microaneurysm-like characteristics using distinct image-processing principles.
- E. No preprocessing: The framework formally includes a “No preprocessing” operation for candidate extractors applied directly to original images.
- Candidate extraction seeks objects in the image showing microaneurysm-like characteristics.
- One extractor finds sufficiently small dark patterns on the green channel using grayscale diameter closing and a double threshold.
- A vessel map, matched filter, fixed threshold, and region growing are used in another candidate-extraction procedure.
- Another extractor detects circular objects using a circular Hough transformation.
D. Zhang et al. [18]
The framework builds ensembles from preprocessing–candidate-extractor pairs, fuses nearby detections by voting and centroids, and selects the best-performing combination using CPM.
- Each ensemble member sequentially applies one preprocessing method and one candidate extractor to generate microaneurysm candidates.
- Nearby candidates from different ensemble members are grouped within distance r, and each group contributes its centroid to the fused candidate set.
- The framework evaluates candidate ensembles against ground truth using distance-based true-positive and false-positive decisions.
- With M = N = 5, the system contains 25 preprocessing–extractor pairs and 2^25 possible ensemble combinations.Because evaluating every combination is resource-demanding, simulated annealing searches for the final ensemble.
- CPM is the average sensitivity at seven predefined false-positive-per-image rates and serves as the ensemble-selection energy function.The selected ensemble Ebest is the highest-performing configuration found by the search.
- During deployment, the selected ensemble fuses detections from its component pairs and thresholds candidate confidence values for final microaneurysm decisions.
V. METHODOLOGY
The study evaluates the proposed approach for both microaneurysm detection and diabetic retinopathy grading.
- The evaluation covers microaneurysm detection and diabetic retinopathy grading.
A. MA detection
Microaneurysm detection is evaluated on the ROC competition dataset, DiaretDB1 2.1, and a private database with differing image properties and lesion distributions.
- The proposed detector is evaluated on the ROC competition, DiaretDB1 2.1, and a private database.
- The ROC database contains 50 training and 50 test images with varied resolutions, 45° field of view, and JPEG compression.Average microaneurysm counts are 6.72 for training images and 6.86 for test images.
- DiaretDB1 2.1 contains 28 losslessly compressed training images and 61 test images at 1500 × 1152 resolution and 50° field of view.Average microaneurysm counts are 4.34 for training images and 3.91 for test images.
- The private database contains 60 losslessly compressed images at 3072 × 2048 resolution and 45° field of view.Average microaneurysm counts are 8.67 for training images and 8.87 for test images.
4) Testing:
Testing measures detection and image-level diabetic retinopathy classification using thresholded confidence values, with Messidor providing 1200 graded images and an independent ROC training set.
- FROC curves report sensitivity against average false positives per image for each evaluated microaneurysm-detection database.Detector outputs are thresholded using candidate confidence values.
- On Messidor, images are classified as diseased for grades R1–R3 and healthy for grade R0 using the presence or absence of detected microaneurysms.
- The algorithm is trained on the ROC dataset because Messidor provides no training set, creating a strong handicap relative to within-database training.
- Messidor contains 1200 losslessly compressed images with 45° field of view and resolutions ranging from 440 × 960 to 2304 × 1536.
- Messidor grading includes 540 R0, 153 R1, 247 R2, and 260 R3 images, corresponding to increasing diabetic-retinopathy severity.
- Sensitivity, specificity, accuracy, per-grade recognition, and empirical and fitted AUC are measured on Messidor after thresholding candidate confidence values.
VI. RESULTS
The proposed ensemble combines preprocessing and candidate-extractor components and is evaluated for microaneurysm detection across multiple datasets. It leads the ROC competition and achieves higher CPM and partial AUC than individual algorithms.
- The ensemble includes selected preprocessing-method and candidate-extractor pairs for three datasets, organized by preprocessing rows and extractor columns.
- The proposed ensemble is the current leader in the ROC competition.Its performance is shown using a FROC curve on the ROC dataset.
- The ensemble achieves higher CPM and partial AUC than individual algorithms in the ROC competition.
- FROC curves report the ensemble’s performance on the DiaretDB1 v2.1 dataset.The corresponding figure is identified as Figure 4.
B. DR grading
The ensemble performs well for diabetic retinopathy grading, with performance varying by disease severity and threshold. On Messidor, it achieves a fitted ROC AUC of 0.90 ± 0.01, while serious cases can still be misclassified.
- DR grading: 0.90 ± 0.01 fitted ROC AUC was achieved on the Messidor dataset for the detector.The empirical AUC was 0.875.
- Discussion: CLAHE can improve MA-background separation in difficult images, while vessel removal and inpainting can cause a true MA to be missed.Using different preprocessing methods with candidate extractors creates diversity within the ensemble.
- Discussion: The ensemble outperforms current individual approaches and maintains stable performance across datasets despite variation in ensemble members.On the ROC database, its best competition score occurred near the human-expert level of approximately 1 false positive per image.
- DR grading: Recognition probability increases with DR severity at threshold levels where sensitivity is below 1.0.The section reports different recognition performance across grades R0, R1, R2, and R3.
- DR grading: 96% sensitivity and 51% specificity were achieved at threshold 0.8, whereas threshold 0.9 produced 76% sensitivity and 88% specificity.Threshold 0.9 gave the highest accuracy and was closest to the BDA recommendation.
- Discussion: Comparison with other screening systems is difficult because studies use different datasets and different proportions of DR cases.Reported comparison values include 0.86, 0.81, and 0.89 AUCs under differing case distributions.
- Discussion: The system still misclassifies some severe DR stages and would need additional DR-specific lesions, such as exudates, for improved grading.The MA detector is presented as a main component rather than a complete clinical screening system.
VIII. CONCLUSION
The paper concludes that a modular ensemble combining preprocessing methods and candidate extractors achieves strong microaneurysm detection and promising diabetic retinopathy grading. On Messidor, it reached an AUC of 0.90 ± 0.01, while a complete screening system would require additional components.
- VIII. CONCLUSION: The proposed modular ensemble combines preprocessing-method and candidate-extractor pairs, with a search algorithm selecting an optimal combination.The modular design permits future expansion by adding preprocessing methods and candidate extractors.
- VIII. CONCLUSION: 0.90 ± 0.01 AUC was achieved for diabetic retinopathy grading on 1,200 Messidor images.The grading results are described as competitive with previously reported results on other databases.
- VIII. CONCLUSION: The detector achieved first position in an open online microaneurysm-detection challenge.The conclusion characterizes its efficiency in that challenge as high.
- VIII. CONCLUSION: A proper screening system should include other components expected to increase performance beyond the proposed detector.The conclusion presents the current grading results as promising rather than complete.