Machine Learning Assisted Quantification of New Vessels on the Disc in Proliferative Diabetic Retinopathy

Abstract

Purpose : The Early Treatment Diabetic Retinopathy Study (ETDRS) severity scale has been the gold-standard for grading diabetic retinopathy (DR) in clinical research for decades. Graders determine ETDRS severity from fundus photographs by comparing lesions to standard photographs. Grading of new vessels on the disc (NVD) is based on the area involved by the new vessels compared to standard 10A (Figure 1) which has NVD equal to ⅓disc area (DA) or 0.85mm². Then, an eye is categorized as <⅓ or ≥⅓DA. However, this binary classification may be insufficient to measure regression of NVD as a potential outcome for therapeutic studies. Our goal is to use a machine learning (ML) tool to quantitatively analyze images

Methods : Fundus photographs from 66 eyes with NVD were included. Images were calibrated using disc to macula distance. Of these, 40 images were randomly selected to train a classifier with Trainable Weka Segmentation, a Fiji plugin which combines ML algorithms to perform pixel-based segmentation. The classifier was applied to the remaining images, and a segmented mask was generated for each. Artifactually-segmented pixels outside the area of NVD were not excluded in the current version of our classifier. NVD area (mm²) and vascular length (mm) were documented for each eye. All images were independently reviewed by two graders and categorized as NVD <⅓ or ≥⅓DA.

Results : Graders categorized 31 (46.9%) eyes as <⅓DA and 35 (53.1%) eyes as ≥⅓DA. There was equal distribution of both categories in training and validation images. In eyes categorized as NVD <⅓DA, the mean area of NVD with ML quantification was 0.099mm² (SEM=0.025), compared to 0.501mm² (0.159) in eyes categorized as ≥⅓DA (p=0.02) Mean vascular length of NVD was 0.010mm (0.003) and 0.048mm (0.016) in both categories respectively (p=0.026).

Conclusions : There is strong correlation between ML quantification of NVD and subjective binary classification. Quantification of NVD using ML provides a finer measurement and ability to assess change over time. Employing this tool in longitudinal data will help describe the natural history of change in area and length of NVD over time.

This is a 2021 ARVO Annual Meeting abstract.

 

Figure 1: ETDRS standard 10A. Scale bar=1mm length.

 

Figure 2A: Image with NVD <⅓DA. 2B: Segmented image with NVD highlighted in yellow (0.090 mm2, 0.008 mm). 2C: Image with NVD ≥⅓DA. 2D: Segmented image with NVD highlighted (0.329 mm2, 0.026 mm). Scale bars=1mm length.