Self-supervised iterative refinement learning for macular OCT volumetric data classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Macular optical coherence tomography (OCT) is crucial for diagnosing retinal diseases.
• Volumetric OCT image classification often relies on 2D B-scan analysis.
• Existing methods may face challenges in accurately classifying complex OCT volumes.

Purpose of the Study:

• To introduce a novel self-supervised iterative refinement learning (SIRL) pipeline.
• To improve the performance of B-scan classification-based macular OCT volumetric image classification algorithms.
• To enhance the accuracy and reliability of OCT-based diagnostic tools.

Main Methods:

• SIRL employs a repetitive training-sieving-relabeling process.
• Initial labels are derived from volume labels for 2D B-scans.
• Iterative refinement updates 2D image labels based on network predictions.

Main Results:

• SIRL achieved high performance on both clinical and public datasets.
• On a clinical dataset, sensitivity, specificity, and accuracy were 89.74%, 94.87%, and 93.18%.
• On a public dataset, corresponding metrics reached 98.22%, 90.43%, and 95.88%.

Conclusions:

• SIRL effectively improves macular OCT volumetric image classification.
• The proposed method enhances algorithms that use B-scan classification.
• SIRL demonstrates significant potential for advancing OCT image analysis in clinical practice.