3DTINC: Time-Equivariant Non-Contrastive Learning for Predicting Disease Progression From Longitudinal OCTs.

Area of Science:

• Medical imaging analysis
• Deep learning in healthcare
• Ophthalmology research

Background:

• Self-supervised learning (SSL) enhances deep learning models but contrastive methods struggle with 3D medical images due to batch size and augmentation limitations.
• Current SSL techniques are often impractical for 3D optical coherence tomography (OCT) data, hindering its application in medical image analysis.

Purpose of the Study:

• To introduce a novel longitudinal self-supervised learning (SSL) method, 3DTINC, for 3D OCT volumes.
• To develop perturbation-invariant feature learning specifically tailored for OCT data.
• To implicitly learn temporal information from longitudinal scans for disease progression prediction.

Main Methods:

• Proposed 3DTINC, a non-contrastive SSL method for 3D OCT volumes.
• Introduced OCT-specific augmentations and a novel non-contrastive similarity loss term.
• Leveraged intra-patient scans from different time points to learn temporal features.

Main Results:

• 3DTINC effectively learns perturbation-invariant features from 3D OCT data.
• The learned temporal information is crucial for predicting retinal disease progression.
• Models pretrained with 3DTINC demonstrated strong performance in predicting wet age-related macular degeneration (AMD) conversion.

Conclusions:

• Each component of 3DTINC is vital for learning meaningful representations from longitudinal volumetric scans.
• The method provides a robust framework for analyzing temporal changes in retinal OCT.
• 3DTINC advances the prediction of disease progression in conditions like AMD using medical imaging.