MDTNet

MDTNet automates analysis of retinal arterio-venous crossing patterns using deep learning to grade arteriolosclerosis for cardiovascular risk stratification.

Key Features:

• Automated Segmentation and Classification: Employs segmentation and classification models to identify vessels in retinal images and label them as arteries or veins to locate candidate arterio-venous crossing points.
• Crossing Point Validation: Uses a classification model to confirm true arterio-venous crossing points, reporting precision and recall rates of 96.3%.
• Severity Grading: Classifies severity of vessel crossings according to Scheie's classification, achieving a kappa value of 0.85 with retina specialists and 92% accuracy.
• Multi-Diagnosis Team Network (MDTNet) Architecture: Combines sub-models with diverse structures or loss functions into an ensemble to address label ambiguity and imbalanced data distribution.
• Explainability and Reproducibility: Replicates ophthalmologist diagnostic steps without relying on subjective feature extraction to provide a transparent and reproducible grading process for arteriolosclerosis.

Scientific Applications:

• Cardiovascular Risk Assessment: Automates and standardizes grading of arteriolosclerosis from retinal images to support cardiovascular risk stratification.
• Clinical Research: Provides objective, reproducible measures of retinal vascular health for studies by ophthalmologists and researchers.
• Early Detection and Stratification: Facilitates early detection and stratification of cardiovascular risk through automated analysis of retinal vascular indicators.

Methodology:

The workflow comprises three explicitly stated stages—vessel segmentation and labeling, crossing point validation, and severity grading—implemented with deep learning models and an ensemble of diverse sub-models with varied structures or loss functions.