Hybrid AI model achieves major leap in detecting eye lesions from retinal images

Researchers have developed a hybrid machine-learning system that significantly outperforms conventional deep-learning approaches at identifying and mapping choroidal nevi—benign pigmented spots in the eye that occasionally transform into melanoma. The breakthrough comes as a response to the persistent challenge of reliably detecting these lesions in fundus images, the colour photographs taken during eye examinations.

The clinical challenge#

Early detection of choroidal nevi is essential, as melanoma arising from these lesions has serious consequences if diagnosis is delayed. Yet clinicians without specialized training often struggle to distinguish them in standard imaging, and the existing datasets used to train AI systems are plagued by poor resolution and inconsistent annotations. These limitations have hampered the development of trustworthy automated diagnostic tools.

How the hybrid approach works#

Rather than relying solely on U-Net, a popular deep-learning architecture that demands large quantities of annotated training data, the researchers combined mathematical and clustering-based segmentation techniques with insights from neural network design. Traditional mathematical methods offer high accuracy but typically require extensive manual effort to optimize. By fusing these two paradigms, the team created a system that preserves the precision of classical approaches while inheriting the scalability benefits of modern AI.

Performance gains#

The hybrid model was tested on high-resolution (1024×1024 pixel) fundus images and achieved a Dice coefficient of 89.7% and an IoU (Intersection over Union) of 80.01%—metrics that measure how accurately the algorithm identifies lesion boundaries. By contrast, Attention U-Net, a leading deep-learning baseline, reached only 51.3% and 34.2% on the same measures. The new approach also proved more robust when applied to external datasets, suggesting it generalizes better across diverse imaging conditions.

Broader applications#

This work is part of a larger initiative to create clinical decision-support software for diagnosing choroidal nevi. Potential downstream uses include automated lesion annotation systems that could accelerate diagnosis and monitoring workflows in ophthalmology, ultimately improving patient outcomes through faster, more accurate assessment.