Abstract
Purpose: :
To establish a reference for the interpretation by specialist clinicians of non-stereoscopic optic nerve head (ONH) images for signs of glaucomatous damage.
Methods: :
Ten specialist clinicians from Canada and the UK took part. ONH images (n=100) (Topcon TRC-50EX, Japan, 24 bit colour, 350x410 resolution, 20º field of view) were obtained from patients of Manchester Royal Eye Hospital with manifest or suspect glaucoma. Clinicians were presented with images in random order on a computer monitor for up to 60s, and time was unlimited for a forced selection from 5 categories (definitely healthy, probably healthy, not sure, probably damaged, definitely damaged). Prior to the session, clinicians were informed of the proportion of images with corresponding visual field (VF) loss, and instructed to make decisions based on ONH damage rather than likelihood of VF loss. No feedback was given during the session.In order to allow ROC analysis for detection of VF loss, the images were split into 2 groups based on 4 consecutive VF records: VF negative (VFN) (n=80) (Pattern Standard Deviation (PSD) better than 2.0dB, Mean Deviation (MD) better than -1.5dB, and fellow eye MD within 1.0dB of the imaged eye and also better than -1.5dB), and VF positive (VFP) (n=20) (MD between -2.5dB and -10.0dB and PSD between 3.0dB and 15.0dB). The two groups were matched for image quality as graded by a masked observer. Images from each group were randomly repeated (30% of VFN, 10% of VFP).
Results: :
Group responses were combined to produce an ROC curve for best available performance in detection of VF loss (area under curve=0.87), sensitivity at 90% specificity was 60%. For individual clinicians, area under ROC curves for detection of VF loss ranged from 0.71 to 0.89. Between-clinician rank correlation in image grading ranged from R=0.47 to R=0.80. There were large and statistically significant differences in the response criteria adopted by clinicians (Friedman test, p<0.001). Clinician's median response time per image ranged from 4 to 16s, 95% of responses being made within 30s.
Conclusions: :
Whilst not providing a true index of in-vivo diagnostic ability of clinicians (absence of diagnostic clues such as stereo depth and between-eye comparison), the results provide a valid reference for comparing results from primary care clinicians and trainees on the same set of images. The software and dataset are freely available via anonymous ftp from www.opticdisc.org.
Keywords: imaging/image analysis: clinical • optic disc • clinical (human) or epidemiologic studies: systems/equipment/techniques