Development and evaluation of customized software to automatically align macula and optic disc centered scanning laser ophthalmoscope fundus images.

In ophthalmology, the angle between the center of the optic nerve head and the center of sharpest vision (foveola) is a posterior fundus landmark parameter of the retina of the human eye. Together with the optic disc-fovea distance, it characterizes the position of the optic nerve head in relationship to the foveola. The optic disc-fovea angle markedly influences the regional distribution of retinal layer thickness patterns, specifically the retinal nerve fiber layer thickness measured at the optic disc. Thus, the optic disc-fovea angle needs to be determined and routinely taken into account in morphological glaucoma diagnosis and in the assessment of structure-function relationship in optic nerve diseases. However, despite the urgency of this information, currently the optic disc-fovea line and its angle are routinely not measured. Obtaining it post-measurement requires manual registration of the macula and optic disc optical coherence tomography (OCT) imaging data. OCT manufacturer-delivered software does not provide automated image registration. Therefore researchers are forced to manually perform the alignment over different scanning regions. To fill this gap, we provide two software packages which can be applied to routinely acquired clinical OCT data to automatically align macula and optic disc images. In this work, we introduce and comparatively evaluate two separate software packages (BloodVesselReg and OCTFundusReg) to automatically align macula and optic disc centered OCT volume scans based on their respective scanning laser ophthalmoscope (SLO) fundus images. BloodVesselReg implements an image registration and mosaicing algorithm based on retinal blood vessels. OCTFundusReg optimizes a general-purpose image registration toolkit to operate on SLO images. Both methods were independently developed by different subgroups of authors of this study using a training dataset of 18,047 eyes from a population-based study. The methods were tested on a dataset of 3,570 eyes from glaucoma patients, with success/failure assessed by visual inspection and compared to failure reporting of the methods themselves. BloodVesselReg had a slightly higher accuracy (94.7%) than OCTFundusReg (93.9%). Both methods together failed on only 1% of the eyes. BloodVesselReg reported 165 out of its 190 failures. OCTFundusReg provides a continuous failureAlert parameter which resulted in an area under the receiver operating characteristics curve (AUC) of 0.91 from a logistic regression model. When including the difference of fitting related parameters between the two methods, the AUC improved to 0.95. Both methods had success rates of over 90% when applied in isolation to a clinical testing dataset. When applying them together, the rate of at least one of the method succeeding was 99%. The methods are highly promising for applications under real-world clinical conditions and might help to facilitate disease detection and monitoring over time.