Related Experiment Video
Updated: Sep 4, 2025
12:28
Quantification of Vascular Parameters in Whole Mount Retinas of Mice with Non-Proliferative and Proliferative Retinopathies
Published on: March 12, 2022
3.7K
AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline
Yukun Zhou1,2,3, Siegfried K Wagner2, Mark A Chia2
1Centre for Medical Image Computing, University College London, London, UK.
Translational Vision Science & Technology
|July 14, 2022
Summary
The AutoMorph deep learning pipeline accurately analyzes retinal vascular morphology from fundus images. This open-source tool enhances research into eye and systemic diseases.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Retinal vascular morphology analysis is crucial for diagnosing ophthalmic and systemic diseases.
- Automated tools are needed to efficiently and comprehensively analyze these complex features.
- Deep learning offers potential for accurate automated analysis of fundus photographs.
Purpose of the Study:
- To externally validate the AutoMorph deep learning pipeline for automated retinal vascular morphology analysis.
- To assess the performance of AutoMorph's individual modules on independent datasets.
- To confirm the pipeline's utility for research in ophthalmic and systemic diseases.
Main Methods:
- AutoMorph comprises modules for image preprocessing, quality grading, anatomical segmentation, and feature measurement.
- Deep learning techniques, including EfficientNet-b4, are employed for image grading and segmentation.
- A model ensemble strategy and confidence analysis are used for robust and accurate results.
- External validation was performed on several independent, publicly available datasets.
Main Results:
- The image grading module achieved an F1-score of 0.86 on EyePACS-Q, comparable to state-of-the-art.
- Confidence analysis improved grading accuracy by reducing misclassified images by 76%.
- Vessel segmentation achieved F1-scores ranging from 0.73 to 0.78, and disc segmentation reached 0.94.
- Vascular morphology features showed good to excellent agreement with expert annotations.
Conclusions:
- AutoMorph modules demonstrate robust performance even with domain shifts in external validation data.
- The fully automated pipeline enables detailed, efficient, and comprehensive retinal vascular morphology analysis.
- Public availability of AutoMorph aims to accelerate research in oculomics and related fields.
