Self-supervised based clustering for retinal optical coherence tomography images.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Manual analysis of retinal optical coherence tomography (OCT) images is insufficient for growing demands.
• Self-supervised learning offers a potential solution for automated OCT image analysis.

Purpose of the Study:

• To implement and evaluate a self-supervised learning-based clustering model for classifying retinal OCT images.
• To compare the model's performance against baseline methods and analyze its feature representation.

Main Methods:

• Utilized a public dataset of 83,484 OCT images (CNV, DME, drusen, normal).
• Employed the Semantic Pseudo Labeling for Image Clustering (SPICE) framework for self-supervised learning.
• Performed binary and four-category clustering, t-SNE analysis, and examined cluster centers and attention maps.

Main Results:

• SPICE achieved superior accuracy: 0.886 for binary and 0.846 for four-category classification.
• t-SNE analysis showed distinct clustering of the four retinal image types.
• Model focused on key biomarkers, and cluster centers aligned with expert labels.

Conclusions:

• Self-supervised clustering effectively differentiates retinal disease variations using OCT images.
• The SPICE model demonstrates capability in detecting disease heterogeneity via biomarkers.
• This approach enhances diagnostic potential for conditions like diabetic macular edema and choroidal neovascularization.