Adaptive Stochastic Conjugate Gradient Optimization for Temporal Medical Image Registration

We propose an Adaptive Stochastic Conjugate Gradient (ASCG) optimization algorithm for temporal medical image registration. This method combines the advantages of Conjugate Gradient (CG) method and Adaptive Stochastic Gradient Descent (ASGD) method. The main idea is that the search direction of ASGD is replaced by stochastic approximations of the conjugate gradient of the cost function. In addition, the step size of ASCG is based on the approximation of the Lipschitz constant of the stochastic gradient function. Thus, this algorithm could maintain the good properties of the conjugate gradient method, meanwhile it uses less gradient computation time per iteration and adjusts the step size adaptively as the ASGD method. As a result, this algorithm takes less CPU time than the previous ASGD method. We demonstrate the efficiency of our algorithm on the public available 4D Lung CT data and our clinical Lung/Tumor CT data using the general 4D image registration model. We compare the ASCG with several existing iterative optimization strategies: steepest gradient descent method, conjugate gradient method, Quasi-Newton method (LBFGS) and adaptive stochastic gradient descent method. Our preliminary results indicate that our ASCG algorithm achieves 22% higher accuracy on the POPI dataset and it also performs better than existing methods on other datasets(DIR-Lab dataset and our clinical dataset). Furthermore, we demonstrate that compared with other methods, our ASCG algorithm is more robust to image noises.