Developing automated vascular leakage segmentation in retinal vasculitis patients using different deep learning architectures
Purpose
Retinal vascular leakage is a crucial finding of retinal vasculitis (RV) manifest in fluorescein angiography (FA). Although the degree of retinal vascular leakage is associated with the severity of RV, there is no standard scheme to segment retinal vascular leakage specifically for RV patients. Here, we developed automated segmentation models for retinal vascular leakage in RV patients based on several deep learning architectures.
Methods
A total of 463 FA images from 83 patients diagnosed with RV have been used to develop the deep learning models. The vascular leakage was identified and manually segmented for each FA image. The images were split at a ratio of 60:20:20 for training, validating and testing datasets. The deep learning models were trained on DeeplabV3+, UNet++ and UNet model architectures using training and validating datasets. The models were evaluated on the testing dataset, and the model with the highest pixel-similarity dice score was selected as the best performing.
Results
Dice scores on the test dataset ranged from 0.5302 to 0.6279. The UNet++ model architecture was the best model for retinal vascular leakage with a dice score of 0.6279 (95% confidence interval 0.5584-0.6974).
Conclusion
We developed a deep learning detection and segmentation model for retinal vascular leakage in RV patients. Although a higher dice score would be required for clinical applications, deep learning is a promising tool for automated quantification of leakage in RV.
Conflict of interest
No
Authors 1
Last name
DHIRACHAIKULPANICH
Initials of first name(s)
D
Department
Department of Eye & Vision Sciences, University of Liverpool
City
Liverpool
Country
United Kingdom
Authors 2
Last name
Xie
Initials of first name(s)
J
Department
Department of Eye & Vision Sciences, University of Liverpool
City
Liverpool
Country
United Kingdom
Authors 3
Last name
Chen
Initials of first name(s)
X
Department
Xiamen Eye Center of Xiamen University
City
Xiamen
Country
China
Authors 4
Last name
Li
Initials of first name(s)
X
Department
Xiamen Eye Center of Xiamen University
City
Xiamen
Country
China
Authors 5
Last name
Madhusudhan
Initials of first name(s)
S
Department
St Paul's Eye Unit, Liverpool University Hospitals NHS Foundation Trust
City
Liverpool
Country
United Kingdom
Authors 6
Last name
Zheng
Initials of first name(s)
Y
Department
Department of Eye & Vision Sciences, University of Liverpool
City
Liverpool
Country
United Kingdom
Authors 7
Last name
Beare
Initials of first name(s)
NAV
Department
Department of Eye & Vision Sciences, University of Liverpool
City
Liverpool
Country
United Kingdom
This website uses cookies to ensure you get the best experience on our website.
Learn more