Automated image quality assessment and landmark localisation in ultra-widefield retinal images

Retinal imaging allows assessment of ocular health, and is important for the detection and monitoring of retinal diseases. One such device for imaging the retina is the ultra-wide field of view scanning laser ophthalmoscope (UWFoV-SLO), manufactured by Optos plc, which is capable of imaging up to 200 degrees of the retina in a single scan. This thesis details work in two areas, both relating to the image processing of UWFoV-SLO images. The first is the investigation of automated image quality assessment algorithms for UWFoV-SLO images. A novel image quality metric (referred to as the GVC metric) was developed, which is based on a textural measure of image patches that contain retinal vasculature. This metric is demonstrated to correlate with the assessment of image quality as graded by experts. The GVC metric was applied to images captured at a UWFoV-SLO manufacturing facility, and timeseries analysis of this data discovered a number of potential system modifications that could improve the quality of manufactured devices. These included the correction of a software bug that was impairing image quality, and to install all manufactured devices with a detector that captured higher quality images. These observations were further confirmed with forced-choice preference testing with expert assessors of image quality. The second investigation concerns the automated localisation of two retinal landmarks (the optic disc and fovea) in UWFoV-SLO images with convolutional neural networks. On a test set of 1485 images, optic disc localisation accuracy of 96.70% less than 1 optic disc radius from the ground truth was achieved, and 93.27% of fovea predictions were less than 1 optic disc radius from the ground truth. It is also shown that the laterality of the image (whether it is of the left or right eye) can be reliably inferred from the landmark coordinates. Finally, three methods to automatically detect unreliable landmark predictions are presented. These are based on prior knowledge of the spatial distribution of landmarks --- two methods employ Gaussian mixture models, and the third applies a threshold to the Euclidean distance between the predicted landmark coordinates.