Efficient Vision Transformers for Ophthalmic Images Classification: A Comparative Study of Supervised,...

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Medical imaging, particularly in ophthalmology, faces challenges with high annotation costs, limiting the development of AI diagnostic tools.
• Supervised learning models require extensive labeled data, which is scarce and expensive to obtain in specialized fields like ophthalmic imaging.
• Developing robust diagnostic systems necessitates methods that can perform effectively even with limited labeled datasets.

Purpose of the Study:

• To investigate the integration of supervised, semi-supervised, and unsupervised learning strategies for classifying ophthalmic images under label-scarce conditions.
• To enhance diagnostic performance in ophthalmology by leveraging minimal labeled data and robust feature representations.
• To evaluate the effectiveness of different transformer architectures and learning strategies in improving AI-driven ophthalmic diagnostics.

Main Methods:

• Utilized a dataset of 18,767 multimodal ophthalmic images (1,877 labeled, 16,890 unlabeled).
• Employed supervised learning with ViT-Base, DeiT-Base, and MaxViT-L transformer architectures.
• Implemented semi-supervised learning via pseudo-labeling (confidence threshold ≥ 0.98) and unsupervised learning using SimCLR-based contrastive learning and K-means clustering.

Main Results:

• In supervised learning, ViT-Base achieved 92.47% accuracy. After semi-supervised pseudo-labeling, MaxViT-L reached 97.49% accuracy and 0.9982 AUC.
• Unsupervised contrastive learning with MaxViT-L improved feature clustering (Silhouette Score: 0.556, DBI: 0.541).
• MaxViT-L demonstrated superior performance on an external validation set, offering a favorable trade-off between accuracy and generalization despite higher computational complexity.

Conclusions:

• MaxViT-L, particularly in semi- and unsupervised settings, provides a strong balance between diagnostic performance and model generalization for ophthalmic image classification.
• The integrated approach effectively minimizes reliance on expert annotations, paving the way for scalable and automated ophthalmic diagnostic solutions.
• This study highlights the potential of advanced machine learning techniques to overcome data scarcity challenges in medical AI, improving diagnostic capabilities in ophthalmology.