U-Net Improves Multi-Lesion Detection in Diabetic Retinopathy Fundus Images
Diabetic retinopathy remains a major cause of vision impairment, and the clinical burden begins with a familiar bottleneck: retinal lesions must be recognized early, accurately, and fast enough to fit into screening workflows. Microaneurysms, haemorrhages, soft exudates, and hard exudates are not interchangeable findings. Each reflects a different stage or pattern of retinal injury, and each can alter the urgency of referral and follow-up. The study positioned diabetic retinopathy as an imaging problem well suited to automated analysis because manual grading is inherently subjective and time-consuming, especially when subtle lesions are present in crowded fundus images.
A unified deep learning framework built around U-Net addressed the challenge of simultaneous multi-lesion detection rather than isolated lesion spotting. The architecture matters because U-Net’s encoder-decoder design preserves fine structural detail while reconstructing pixel-level masks, a feature that is especially important for small lesions such as microaneurysms. In retinal imaging, lesion boundaries are often faint, irregular, and easy to miss when the model sacrifices spatial precision for global pattern recognition. A segmentation framework that can keep those details in view is better aligned with the actual clinical task.
Training Strategy and the Role of Adaptive Batch Scaling
The model was paired with an Adaptive Batch Scaling strategy, or ABS, in which batch size changed dynamically across training epochs according to validation Dice score. This is a practical design choice rather than a cosmetic one. Batch size affects gradient stability, convergence behavior, and the model’s ability to improve segmentation quality without drifting into overfitting or underfitting. By using validation performance to guide training configuration, the framework attempted to tune optimization to the lesion-segmentation problem itself rather than forcing a fixed training regime across all stages.
The study compared U-Net with other deep learning segmentation models, including Mask Region-based Convolutional Neural Network, YOLOv8, and Faster Region-based Convolutional Neural Network architectures. The comparison is clinically relevant because retinal screening systems do not compete on novelty alone. They compete on whether they can detect small lesions consistently across real-world image variation and without excessive computational burden. A single-pass system that identifies multiple lesion categories can reduce processing overhead while still supporting screening decisions.
Performance Across DDR and IDRiD Datasets
Validation used two publicly available datasets, the Indian Diabetic Retinopathy Image Dataset (IDRiD) and the Diabetic Retinopathy Detection Dataset (DDR). Training was performed on DDR, and the model was then tested separately on DDR and IDRiD to evaluate both intra-dataset performance and cross-dataset generalization. That choice is important because a model may perform well on the data distribution it has already learned and lose accuracy when exposed to a different image set, capture style, or lesion pattern. In retinal AI, transferability often determines whether a tool remains useful beyond the lab.
Across the evaluation metrics reported, U-Net performed best. It achieved an accuracy of 0.925, a Dice score of 0.771, an Intersection over Union of 0.748, and a mean average precision of 0.647. Those numbers indicate reasonably strong pixel-level overlap and lesion localization in a segmentation setting where even modest gains can matter clinically. Dice and IoU are especially informative here because they reflect how closely the predicted lesion masks matched the reference annotations, not merely whether the algorithm labeled an image as diseased.
A single-pass segmentation system can change how diabetic eye screening is operationalized.
Why Multi-Lesion Segmentation Matters at the Bedside
The biological logic behind the system is straightforward. Diabetic microvascular injury does not present as one uniform lesion type, and retinal screening is most useful when it captures the full spectrum of visible damage in a single read. Microaneurysms may be the earliest sign of capillary wall weakness, while haemorrhages and exudates suggest more advanced vascular leakage and lipid deposition. A model that detects several lesion classes together gives a more complete picture of retinal injury and can support triage decisions more efficiently than a model built around a single abnormality.
The study’s emphasis on semantic segmentation rather than broad image classification is clinically meaningful. Classification can say whether disease is present, but segmentation can show where and how much. In practice, lesion burden and distribution often shape interpretation, particularly when clinicians need to judge disease extent, monitor progression, or decide whether a patient needs specialist review. A segmentation tool that preserves lesion detail can therefore support both screening and documentation.
Limits of the Dataset-Driven Design
The framework was evaluated using publicly available DDR and IDRiD datasets, and no additional datasets were introduced. That keeps the work focused and reproducible, but it also means performance remains tied to the properties of those datasets. Real-world screening programs may encounter different camera systems, lighting conditions, patient populations, or lesion prevalence patterns. Cross-dataset testing helps address that concern, yet it does not fully reproduce the diversity of clinical practice. The reported results are encouraging, but broader external validation would be needed before the approach could be treated as a routine screening tool.
No funding supported the research, and the authors declared no competing interests. The study therefore reads as a technical demonstration of what a carefully tuned segmentation model can do when the target is multi-lesion retinal analysis. Its central contribution is not simply that deep learning can detect diabetic retinopathy, but that one framework can localize several lesion types at once with useful performance and lower computational load than separate lesion-specific systems.
For diabetic eye care, the remaining question is not whether automated segmentation can work on benchmark datasets. The harder question is how reliably it will hold up when faced with the messy heterogeneity of clinical screening programs, where image quality, lesion subtlety, and referral thresholds all vary at the point of care.