Pretraining of 3D image segmentation models for retinal OCT using denoising-based self-supervised learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning automates retinal biomarker segmentation in Optical Coherence Tomography (OCT) for clinical research and patient monitoring.
• Current methods, like U-Nets, require supervised training and extensive, expert-annotated datasets, which are difficult and costly to obtain.

Purpose of the Study:

• To investigate the effectiveness of 3D self-supervised learning (SSL) using image restoration techniques for pretraining networks.
• To enhance the performance of fluid segmentation in retinal OCTs and reduce the dependency on manual annotations.

Main Methods:

• Explored two SSL methods: image restoration and denoising, for pretraining 3D networks on a large 3D OCT dataset.
• Evaluated pretrained network weights by fine-tuning them on two distinct fluid segmentation datasets with varying amounts of training data.

Main Results:

• Both SSL methods significantly improved fluid segmentation performance compared to traditional supervised approaches.
• Denoising-based SSL demonstrated superior results on both fluid segmentation datasets and achieved faster pretraining durations.
• The SSL approach enabled either a reduction in required annotations or an increase in segmentation accuracy.

Conclusions:

• 3D self-supervised learning, particularly denoising-based methods, offers a viable and efficient strategy for improving fluid segmentation in retinal OCTs.
• This technique can substantially lower the annotation burden and enhance the precision of automated segmentation in ophthalmology.
• SSL provides a promising avenue for advancing AI-driven clinical research and patient monitoring in retinal imaging.