Transfer learning and vision transformer for the automatic diagnosis of cataracts in ophthalmological images

Cataracts continue to be the leading cause of preventable blindness worldwide and represent a significant public health challenge, particularly in rural and underserved regions where access to ophthalmology specialists and diagnostic infrastructure is limited. Early detection plays a crucial role in preventing visual impairment and improving treatment outcomes; however, large-scale screening programs are often constrained by the availability of trained professionals and specialized equipment. The purpose of this research was to develop and evaluate an automated cataract detection system based on deep learning using retinal fundus images in order to support early screening and improve accessibility to ophthalmological diagnosis.

The proposed methodology followed an experimental quantitative approach that included dataset preparation, image preprocessing, model training, and performance evaluation. A labeled subset of 2,658 retinal fundus images extracted from the ODIR-5K dataset was used as the primary data source. The images underwent preprocessing procedures including normalization and noise reduction, followed by data augmentation techniques such as random rotations (±10°), scaling (90%-110%), and brightness and contrast adjustments. These transformations allowed the creation of a balanced dataset of 4,840 images, enhancing model generalization and reducing overfitting. Six deep neural network architectures were trained and evaluated: ResNet152, EfficientNet-v2S, Inception v3, MobileNet v3, DenseNet201, and Vision Transformer (ViT). Transfer learning with ImageNet pre-trained weights was applied together with selective fine-tuning of deeper layers and optimization using the Adam algorithm combined with a Cosine Annealing learning rate scheduler.

The results obtained indicate that ResNet152 is the bestperforming architecture with an accuracy of 99.10%, precision of 99.72%, sensitivity of 98.46%, and F1 score of 99.08%. It is concluded that deep convolutional neural network architectures, particularly ResNet152, provide highly effective performance for automated cataract detection from retinal fundus images.

The proposed system demonstrates strong potential as a clinical decision-support tool for large-scale screening programs, especially in resource-limited settings, as it can facilitate early diagnosis, improve access to ophthalmological care, and reduce the diagnostic workload of specialized medical personnel.