Deep Learning Model Shows Proficiency for Retinal Artery, Vein Segmentation

A deep learning algorithm, the multiscale guided attention network for retinal artery and vein segmentation and classification (MSGANet-RAV), can segment veins and arteries from retinal and optic disc images, according to research published in the Journal of Optometry. This method may detect sight-threatening conditions such as diabetic retinopathy and glaucoma, according to the report. 

After its development, the MSGANet-RAV algorithm was tested on color optic disc images sourced from the Lion’s Eye Central (LEI-CENTRAL; n=383; patient age range, 16-19 years) database and color fundus images from a publically-available database (AV-DRIVE; n=40; patient age range, 31-86 years). The researchers compared algorithm performance with existing algorithms.

The MSGANet-RAV algorithm outperformed several reference models with respect to retinal blood vessel segmentation and demonstrated a pixel classification accuracy of 93.15%, sensitivity of 92.19%, and specificity of 94.13% for images from the LEI central database. The algorithm demonstrated similar performance when classifying pixels from the AV-DRIVE system with an accuracy of 95.48%, sensitivity of 93.59%, and specificity of 97.27%. For the other algorithms, balanced accuracy ranged between 85.56% and 90.79%, sensitivity between 84.16% and 89.56%, and specificity between 86.96% and 92.02%.

“MSGANet-RAV could be used in automated systems designed to quantitatively assess morphological and functional vascular changes in retinal and optic disc images,” according to the study authors. “The method can be tested in clinical settings for early diagnosis and progression of sight-threatening conditions such as vascular occlusions, glaucoma, and diabetic retinopathy and for automated indirect measurement of [intracranial pressure].”

This research is limited by poor algorithm performance in images obtained from patients with specific vein architectures.

Disclosure: One study author declared affiliations with biotech, pharmaceutical, and/or clinical research organizations. Please see the original reference for a full list of authors’ disclosures.