Change Detection Based on Tensor RPCA for Longitudinal Reitnal Fundus Images

Abstract Change detection of longitudinal fundus images is an important problem in computer aided diagnosis system (CAD). Detecting regions of change in multiple fundus images from the same eye is seldom developed in the literature due to the complication and interpretability. This paper presents a longitudinal change detection framework based on tensor robust principal component analysis (RPCA) for a long retinal fundus image serial. The proposed method chooses an image of the best condition in serial as the background, then models each image as a slice of tensor, utilizes total variation to constraint the temporal continuity of change regions, finally obtains the change regions by Tucker decomposition and alternating direction method of multipliers (ADMM). Comparing with the method based on matrix RPCA, tensor RPCA preserves the original spatial structure of each image, imposes the temporal continuity on change regions and models the background by patches to avoid the little disturbance of blood vessels. Results on a real fundus image serial are presented and show the effectiveness of the proposed algorithm.