MBT: Model-Based Transformer for retinal optical coherence tomography image and video multi-classification.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Retinal disease detection is a critical classification problem in medical imaging.
• Optical coherence tomography (OCT) is a key technology for visualizing retinal structures.
• Transformer models show promise but require performance enhancement for clinical applications.

Purpose of the Study:

• To develop an effective technique for automated retinal disease detection using OCT data.
• To improve the performance of existing transformer models for OCT image and video classification.
• To introduce a novel approach for enhanced feature representation and classification of OCT data.

Main Methods:

• Proposed a Model-Based Transformer (MBT) technique leveraging pre-trained Vision Transformer, Swin Transformer, and Multiscale Vision Transformer models.
• Employed approximate sparse representation for OCT data feature extraction and optimal feature estimation.
• Applied the MBT approach to both OCT image and OCT video classification tasks.

Main Results:

• The MBT method significantly outperformed state-of-the-art deep learning approaches on real-world retinal datasets.
• Achieved superior classification accuracy, precision, recall, F1-score, Kappa, AUC-ROC, and AUC-PR.
• Demonstrated improved performance for Vision Transformer, Swin Transformer, and Multiscale Vision Transformers in OCT analysis.

Conclusions:

• The MBT approach offers a novel method for automated retinal disease detection from OCT data.
• This study highlights the potential of using OCT videos, not just images, for identifying retinal pathologies.
• The findings provide a valuable framework for enhancing deep learning models in ophthalmology research and clinical practice.