RNFLT2Vec: Artifact-corrected representation learning for retinal nerve fiber layer thickness maps.

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Biomedical Data Science

Background:

• Optical coherence tomography (OCT) imaging is vital for glaucoma diagnosis via retinal nerve fiber layer thickness (RNFLT) maps.
• Neural models struggle with individual anatomical variations and artifacts in RNFLT maps, complicating glaucoma feature extraction.
• Accurate RNFLT analysis is crucial for early glaucoma detection and monitoring disease progression.

Purpose of the Study:

• To develop an unsupervised learning framework, RNFLT2Vec, for vectorized feature representation from RNFLT maps.
• To address challenges posed by anatomical variations and artifacts in RNFLT maps for glaucoma-related feature extraction.
• To enhance the accuracy of glaucoma detection and visual field prediction using improved RNFLT representations.

Main Methods:

• Proposed RNFLT2Vec framework for unsupervised learning of RNFLT map features.
• Incorporated an artifact correction component to rectify erroneous RNFLT values.
• Utilized contrastive and consistency learning-based regularization for discriminative representation learning.

Main Results:

• RNFLT2Vec demonstrated superior performance in RNFLT pattern discovery, glaucoma detection, and visual field prediction.
• The artifact correction component effectively produced artifact-free RNFLT map representations.
• Extensive experiments on a large-scale dataset validated the framework's effectiveness.

Conclusions:

• RNFLT2Vec provides a robust method for unsupervised feature learning from RNFLT maps, overcoming common challenges.
• The framework shows significant potential for advancing glaucoma understanding, diagnosis, and patient management.
• This approach facilitates more accurate biomarker identification and prediction of disease outcomes.