Diabetic retinopathy remains one of the leading causes of preventable blindness worldwide, and early detection is crucial for preserving vision. However, initial lesions — microaneurysms, hemorrhages, hard exudates, and soft exudates — are small, low-contrast, and easily missed even by experienced specialists. Traditional automated systems usually handle these four classes with a single multi-class model, but this approach is limited by lesion heterogeneity: they vary drastically in size, color, morphology, and prevalence, favoring common classes over rare ones. Facing this challenge, PRISM-DR emerges, a lesion-specific pipeline that trains one single-class detector per lesion with independent configurations, achieving results that open new perspectives in AI-assisted diagnosis.
PRISM-DR processes raw fundus images by applying region-of-interest cropping, fundus-specific preprocessing, four parallel YOLO detectors (one per lesion), tiling, per-lesion ensembling of five cross-validation folds, and an inter-lesion suppression step that resolves overlaps based on physical lesion size and clinical priority, rather than just model confidence. For each lesion, the best of five YOLO generations is selected and data augmentation is optimized through Bayesian optimization. Trained on the IDRiD dataset with stratified five-fold cross-validation, the system achieves a mAP50 of 0.527 and an F1 of 0.529, with the highest AP50 on hard exudates (0.561). Without fine-tuning, models transfer well when image scale is close to IDRiD, but degrade as field of view and resolution diverge. While these absolute values are modest, they reflect a small training set and a difficult task; nevertheless, treating each lesion as a separate detection problem is a practical alternative to a single multi-class model.
This specialized approach not only improves accuracy on rare lesions but also opens the door to more flexible integrations in real clinical environments. At Q2BSTUDIO, we understand that personalized medicine requires custom software solutions that adapt to the specific needs of each healthcare center. Our experience in aplicaciones a medida (custom applications) allows us to develop modular systems that, like PRISM-DR, treat each subproblem with the optimal tool. Moreover, the scalability of these systems relies on robust cloud infrastructures, whether AWS or Azure, ensuring AI models can be deployed without computational bottlenecks.
From a business perspective, the implementation of specialized detectors like PRISM-DR requires careful orchestration of data, models, and processes. This is where artificial intelligence and AI agents become relevant: not only for detection itself but for automating annotation, validation, and continuous update workflows. At Q2BSTUDIO we integrate AI agents that monitor model performance, suggest retraining, and facilitate clinical auditing, all under cybersecurity protocols that protect sensitive patient information. Our ciberseguridad services ensure any diagnostic system complies with regulations like HIPAA or GDPR, while BI and Power BI solutions visualize performance metrics in interactive dashboards, helping healthcare managers make informed decisions.
The lesson from PRISM-DR is clear: in complex problems with imbalanced classes, specialization outperforms generalization. This philosophy directly transfers to enterprise software development. When a company needs an inventory management system, an e-learning platform, or a virtual assistant, opting for independent, highly optimized modules usually yields better results than a monolithic approach. Q2BSTUDIO applies this principle in every project, combining cloud technologies, artificial intelligence, and data analytics to deliver solutions that evolve with the business. Diabetic retinopathy detection is just one example of how AI, properly segmented and configured, can save lives and optimize resources.
The future of diagnostic imaging lies in modular and specific architectures, where each detector is trained with representative data and intelligent augmentation. PRISM-DR demonstrates that, although absolute results may be modest with limited datasets, the path to excellence is iterative and data-driven. At Q2BSTUDIO we accompany healthcare and technology organizations along that journey, offering everything from prototype conception to production deployment, always with a focus on clinical and business value. Diabetic retinopathy doesn't wait, and neither does software innovation.
