Domain Generalisation for Glaucoma Detection in Retinal Images from Unseen Fundus Cameras.

Out-of-distribution data produced by unseen devices can heavily impact the accuracy of machine learning model in glaucoma detection using retinal fundus cameras. To address this issue, we study multiple domain generalisation methods together with multiple data normalisation methods for glaucoma detection using retinal fundus images. RIMONEv2 and REFUGE, both public labelled glaucoma detection datasets that capture fundus camera device information, were included for analysis. Features were extracted from images using the ResNet101V2 ImageNet-pretrained neural network and classified using a random forest classifier to detect glaucoma. The experiment was conducted using all possible combinations of training and testing camera devices. Images were preprocessed in five different ways using either single or combination of three different preprocessing methods to see their effect on generalisation. In each combination, images were preprocessed using median filtering, input standardisation and multi-image histogram matching. Standardisation of images led to greater accuracy than other two methods in most of the scenarios with an average of 0.85 area under the receiver operator characteristic curve. However, in certain situations, specific combinations of preprocessing techniques lead to significant improvements in accuracy compared to standardisation. The experimental results indicate that our proposed combination of preprocessing methods can aid domain generalisation and improve glaucoma detection in the context of different and unseen retinal fundus camera devices.