Colour normalisation of retinal images

The development of a nationwide eye screening programme for the detection of diabetic retinopathy has generated much interest in automated screening tools. Currently most such systems analyse only intensity information — discarding colour information if it is present. Including colour information in the classification process is not trivial; large natural variations in retinal pigmentation result in colour differences between individuals which tend to mask the more subtle variation between the important lesion types. This study investigated the effectiveness of three colour normalisation algorithms for reducing the background colour variation between subjects. The normalisation methods were tested using a set of colour retinal fundus camera images containing four different lesions which are important in the screening context. Regions of interest were drawn on each image to indicate the different lesion types. The distribution of chromaticity values for each lesion type from each image was plotted, both without normalisation and following application of each of the three normalisation techniques. Histogram specification of the separate colour channels was found to be the most effective normalisation method, increasing the separation between lesion type clusters in chromaticity space and making possible robust use of colour information in the classification process.