Developing a Continuous Severity Scale for Macular Telangiectasia Type 2 Using Deep Learning and Implications for...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning models excel at medical diagnosis but are limited by discrete labels.
• A continuous severity scale for Macular Telangiectasia (MacTel) type 2 could provide more detailed diagnostic information.
• Current diagnostic methods for MacTel may not fully capture disease severity nuances.

Purpose of the Study:

• To develop a novel continuous severity scaling system for MacTel type 2.
• To combine deep learning classification with Uniform Manifold Approximation and Projection (UMAP) for continuous scaling.
• To enhance diagnostic capabilities beyond discrete labels for MacTel severity.

Main Methods:

• A deep learning network was trained on Optical Coherence Tomography (OCT) volumes to learn MacTel severity features.
• Uniform Manifold Approximation and Projection (UMAP) was used to embed these features into a 2-dimensional continuous scale.
• A multiview deep learning classifier was trained on 2003 OCT volumes from 1089 participants.

Main Results:

• The deep learning classifier achieved 63.3% top-1 accuracy on held-out OCT data.
• The UMAP continuous scale demonstrated a strong Spearman rank correlation of 0.84 with a previously established discrete scale.
• The UMAP scale showed substantial agreement (κ = 0.56–0.63) with clinical experts, comparable to interobserver variability.

Conclusions:

• A continuous MacTel severity scale was successfully generated using UMAP embedding without continuous training labels.
• This technique holds potential for application to other diseases, improving diagnosis and understanding of disease progression.
• The developed continuous scale aids in identifying key imaging features related to MacTel pathology.