A comparison of manual and automated methods of measuring conjunctival vessel widths from photographic and digital images

We investigated the application of a fully automated computer algorithm for identifying vessels of the conjunctiva from their scleral surround, and compared measures of vessel width with established methods. Vessel widths at 101 locations (ranging from 20 to 140 mum), from 12 patients, were measured from film and digital images, using a variety of methods, and compared. Widths were measured manually, by semi-automated methods using grey level (densitometric) profiles taken from digital images, and by automated techniques set at different operating levels. Good intra-session repeatibility was obtained using the automated method with an operating sigma value of 3 pixels (16 mum) (mean difference 0.5 mum, 95% CI -8.5 to 9.4 mum) and manual calliper measurements from digitally created photographic slides (mean difference 0.4 mum, -9.3 to 10.1 mum). For comparison with other measures of width, the latter was used as the gold standard. Widths measured from film were slightly larger than those measured directly from digital images, although this effect was small (5 mum) for most vessels. Overall widths measured using the automated method, with a sigma value of 3 pixels, agreed best with the gold standard (inter-method repeatibility; mean difference 1.4 mum, -32.5 to 35.2 mum) although the automated method overestimated small widths (<40 mum) and underestimated larger vessel widths (>40 mum). Automated detection of vessels of the conjunctiva from digital images avoids manual and operator involved measures which are time consuming, and which preclude large patient studies. The resulting data may help in monitoring the vascular response of the conjunctiva to surgical or pharmacological intervention, and in describing vascular changes in response to ocular or systemic disease. The application of this algorithm to the study of retinal vessels is yet to be realised.