Medical Image Retrieval Based On the Parallelization of the Cluster Sampling Algorithm

Cluster sampling algorithm is a scheme for sequential data assimilation developed to handle general non-Gaussian and nonlinear settings. The algorithm relaxes the Gaussian prior assumption widely used in the data assimilation context to approximate the prior distribution obtained by integrating the posterior distribution in previous assimilation cycles. The algorithm can be in general used to solve inverse problems even when linearity or Gaussianity assumptions fail. The cluster sampling algorithm can be used to solve a wide spectrum of problems that requires data inversion such as image retrieval, tomography, weather prediction amongst others. In this paper, we develop parallel cluster sampling algorithms, and show that a multi-chain version is embarrassingly parallel, and can be used efficiently for medical image retrieval amongst other applications. Moreover, we present a detailed complexity analysis of the prposed parallel cluster samplings scheme and discuss their limitations. Numerical experiments are carried out using a synthetic one dimensional example, and a medical image retrieval problem. The experimental results show the accuracy of the cluster sampling algorithm to retrieve the original image from noisy measurements, and uncertain priors. Specifically, the proposed parallel algorithm increases the acceptance rate of the sampler from 44% to 93% with Gaussian proposal kernel, and achieves an improvement of 29% over the optimally-tuned Tikhonov-based solution for image retrieval. The parallel nature of the prposed algorithm makes the it a strong candidate for practical and large scale applications.