Self-supervised patient-specific features learning for OCT image classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning for image classification requires large annotated datasets.
• Manual annotation of optical coherence tomography (OCT) data by ophthalmologists is labor-intensive.
• This limits the application of deep learning in OCT image classification.

Purpose of the Study:

• To propose a self-supervised patient-specific features learning (SSPSF) method.
• To reduce the data annotation requirements for OCT image classification.
• To improve the efficiency and applicability of deep learning in OCT analysis.

Main Methods:

• The SSPSF method involves two phases: self-supervised learning and downstream OCT image classification.
• The self-supervised phase includes tasks to discriminate patient-specific OCT scans and learn patient-invariant features.
• The model learns inherent representations from unlabeled OCT images for effective initialization.

Main Results:

• The SSPSF method achieved high classification accuracies of 97.74% on the RETOUCH dataset and 98.94% on the AI Challenger dataset.
• Experimental results demonstrate the method's effectiveness compared to other OCT image classification techniques.
• The approach successfully utilizes less annotated data.

Conclusions:

• The proposed SSPSF method effectively reduces the need for manual labels in OCT image classification.
• Self-supervised learning provides robust feature representations for OCT data.
• SSPSF offers a promising solution for data-efficient deep learning in ophthalmic image analysis.