Source-linked AI summary
Deep Learning Fundus Image Analysis for Diabetic Retinopathy and Macular Edema Grading
Jaakko Sahlsten, Joel Jaskari, Jyri Kivinen, Lauri Turunen, Esa Jaanio, Kustaa Hietala, Kimmo Kaski
TL;DR
Manual retinal-image screening for diabetic retinopathy and macular edema is labor-intensive, motivating automated grading. The paper develops a deep-learning system evaluated across five screening and clinical grading systems, using relatively few training images and varied resolutions. It reports comparable or better performance with less than one-quarter of the training images used in some prior work, and improved accuracy with high-resolution images.
Problem
Manual retinal-image screening creates a labor-intensive burden as diabetes and screening-image volumes increase.
Method
A deep-learning system was evaluated across five diabetic retinopathy and macular edema classification systems using preprocessing, regularization, different training-set sizes, and image resolutions.
Results
The system achieved comparable or better performance with only a small fraction (<1/4) of training images and improved accuracy with high-resolution images.
Takeaways & Limitations
The results support applying deep learning to accurate diabetic retinopathy and macular edema severity grading on relatively small retinal-image datasets.
Takeaways & Limitations
Messidor comparisons are limited because this study used a different larger Messidor-2 dataset and labels translated from provided Messidor labels were not guaranteed to translate correctly.
Abstract
from arXiv · showhide
Diabetes is a globally prevalent disease that can cause visible microvascular complications such as diabetic retinopathy and macular edema in the human eye retina, the images of which are today used for manual disease screening. This labor-intensive task could greatly benefit from automatic detection using deep learning technique. Here we present a deep learning system that identifies referable diabetic retinopathy comparably or better than presented in the previous studies, although we use only a small fraction of images (<1/4) in training but are aided with higher image resolutions. We also provide novel results for five different screening and clinical grading systems for diabetic retinopathy and macular edema classification, including results for accurately classifying images according to clinical five-grade diabetic retinopathy and four-grade diabetic macular edema scales. These results suggest, that a deep learning system could increase the cost-effectiveness of screening while attaining higher than recommended performance, and that the system could be applied in clinical examinations requiring finer grading.
Introduction
Manual retinal-image screening places a labor-intensive burden on medical experts and healthcare services, motivating automated deep-learning support. This study targets clinically used grading scales while examining performance with fewer training images and varying image resolutions.
- Increasing diabetes prevalence is expected to increase screening images, adding labor-intensive burden for medical experts and healthcare costs.
- Automated systems could support medical experts’ work or serve as a full diagnosis tool.
- Clinical viability requires classifying retinal images according to clinically used diabetic retinopathy and macular edema severity scales.
- Prior experiments addressed the international diabetic retinopathy scale, but no experiments had classified macular changes using the corresponding macular edema scale.
- The study identifies retinopathy across five classification systems, including clinically used PIRC and PIMEC scales.
- The study investigates preprocessing, regularization, training-set size, and image-resolution effects on deep-learning performance.
- The dataset contains 41122 graded retinal color images from 14624 patients, with patient-level separation across training, tuning, and validation sets.
- Images were cropped to tightly contain the fundus, resized to standard input sizes, and divided into training, tuning, and primary validation sets.
Results
Across binary and multiclass diabetic retinopathy and macular edema tasks, the system achieved strong performance, often matching or exceeding prior systems while using far fewer training images. Higher-resolution inputs generally improved results, although performance varied by dataset and comparison.
- 0.987 AUC and 0.940 accuracy outperformed Guan et al.’s 0.9745 AUC and 0.9093 accuracy for NRDR/RDR classification.
- 0.989 AUC was achieved for RDME detection with 2095 × 2095 input images, while specificity was slightly below Gulshan et al. at a similar sensitivity.
- 0.962 macro-AUC, 0.869 accuracy, and 0.910 quadratic weighted kappa were achieved for PIRC classification.
- The model used 28512 images, 24 % of Gulshan et al.’s 118419 images, 37 % of Ting et al.’s 76370, and 1.7% of Krause et al.’s 1662646.
- The system achieved comparable or better performance than state-of-the-art systems while using considerably fewer training images.
- 0.981 macro-AUC, 0.934 accuracy, and 0.856 quadratic-weighted kappa were achieved for PIMEC classification.
- Increasing input resolution from 256 × 256 to 512 × 512 improved binary-task results, with 2095 × 2095 often performing best.
Discussion
The study presents a systematic deep-learning methodology for five diabetic retinopathy and macular edema classification systems, finding strong performance with relatively few training images and generally better results at higher input resolutions. The discussion also identifies computational, dataset, grading, and reproducibility constraints that limit interpretation and deployment.
- Performance and data requirements: The system achieved comparable or better results using less than one-quarter of the training images used in earlier work.The authors attribute performance partly to regularizing preprocessing and dataset features, including grade balance and the experimental setting.
- Dataset and grading limitations: The reported accuracy may reflect a relatively homogeneous population, standardized imaging, grading quality, and an intentionally balanced grade distribution rather than clinical population distributions.The authors note that two-grader reference labels may contain grader biases that reduce generalization.
- Methodological scope: The study evaluates five classification systems and investigates how training-image size affects performance, an assessment the authors describe as previously unconsidered.Classifiers were trained with five input image sizes for each classification system.
- Image resolution: For AUC and macro-AUC, the largest 2095 × 2095-pixel input generally performed best, while quadratic-weighted kappa and accuracy could favor different resolutions under QRDR classification.Under QRDR, the best results used 1024 × 1024 and 2095 × 2095 pixels, respectively.
- Computational trade-offs: Under fixed wall-clock constraints, smaller images may permit broader hyperparameter searches and potentially better results within the same time.This is presented as a practical trade-off between resolution and computational cost.
- Dataset and grading limitations: The system can learn image features unknown or ignored by medical experts because it receives images and grades without diagnostically defining features such as microaneurysms and exudates.These features and their numbers are described as important diabetic retinopathy biomarkers.
- Overall conclusion: The authors report accurate severity grading for diabetic retinopathy and macular edema in a relatively small retinal-image dataset, with accuracy improved by high-resolution, high-quality images.The study also reports that preprocessing and deep-learning code contains proprietary parts and that the primary datasets are not publicly available.
Figures
The figures and tables evaluate binary and multiclass retinal-image classification across input resolutions, datasets, and clinical grading tasks. Results are reported with ROC-based metrics, operating-point measures, and confusion matrices.
- Figure 1: Figure 1 compares ROC curves for referable diabetic retinopathy and macular edema on the primary validation and Messidor sets.Input sizes range from 256 × 256 to 2095 × 2095 pixels, with AUC values shown in the legends.
- Figure 2: Figure 2 shows one-vs-all ROC curves and macro-averages for the best-performing PIRC, PIMEC, and QRDR models.The best models are selected based on macro-AUC, with positive classes and AUC values identified in the legend.
- Tables: The primary-validation tables report classification results for PIRC, QRDR, PIMEC, NRDR/RDR, and NRDME/RDME across varying input image sizes.Binary tasks use sensitivity, specificity, accuracy, and AUC; multiclass tasks additionally use macro-AUC and quadratic-weighted kappa.
- Tables: The Messidor table reports RDR and RDME classification results at a 0.900-sensitivity operating point, including confidence intervals and input-size conditions.Messidor images may be upscaled from 900 × 900 pixels using bicubic interpolation.
- Tables: Confusion matrices organize ground-truth classes by rows and predicted classes by columns for PIRC, PIMEC, and QRDR.PIRC has five classes, PIMEC has four, and QRDR has three, with class labels defined numerically.
Supplementary Information
The supplementary information describes the classifier architecture, optimization across multiple image resolutions, and procedures for selecting and training models.
- Architecture: The baseline architecture is Inception-v3 with its fully connected layer replaced by dropout and softmax-connected layers for diabetic retinopathy grade probabilities.The modified head uses dropout as a regularization technique before the final softmax layer.
- Optimization: Network parameters were fine-tuned with Adam using input sizes of 2095 × 2095, 1024 × 1024, 512 × 512, 299 × 299, and 256 × 256 pixels.Learning rate, dropout rate, and mini-batch size were tuned on the tuning set.
- High-resolution models: The 2095 × 2095 models used mini-batch size 1 and replaced batch normalization with instance normalization because of GPU memory restrictions.These models also used modified training procedures and architecture.
- Model selection: Early stopping used AUC for binary tasks and macro-AUC for multiclass tasks, while the learning rate decayed exponentially during training.The experiments were implemented with Keras and TensorFlow.