Estimating glucose concentration using a scalable sparse mean-shift algorithm

The mean-shift algorithm is a popular approach for clustering and mode location estimation, notably for medical image processing. However, it comes with a high computational cost. In this work, we investigate the possibility of reducing this load using a sparse approximation of the kernel density estimator, based on an incoherence criterion. We derive a sparse scalable mean-shift algorithm that includes a robust formulation of the mean-shift. Our approach is validated in a glucose measurement scenario where we estimate the intensity of a photometric chemical reaction that can be related to an underlying glucose concentration. Our results show that the scalable mean-shift algorithm results in a significant gain in computation time, while staying in the permissible accuracy range for glucose measurements.