Abstract
Purpose:
Our objective was to develop a prototype fully automated detector for capillary abnormalities in ultra-wide-field fluorescein angiograms (UWFA) of branch retinal vein occlusion (BRVO).
Methods:
Four new UWFAs of BRVO were obtained. For each case, three annotations of abnormal capillaries were created by two independent clinicians. Intra- and inter-grader agreement was measured. Composite annotations were created for algorithm validation. An automated algorithm for detection of abnormal capillaries in BRVO was prospectively tested against these composite annotations.
Results:
Near-perfect intra-grader agreement (Fleiss’ kappa range 0.81 - 0.92) and substantial inter-grader agreement (Cohen’s kappa range 0.73 - 0.84) was found among human annotations. Substantial agreement was found between composite annotations and the output of the fully automated detector (Cohen’s kappa for Case 1: 0.59, Case 2: 0.72, Case 3: 0.58, Case 4: 0.69).
Conclusions:
Our fully automated algorithm showed good agreement with human expert annotations.