Abstract
A major problem with Hidden Markov Models (HMMs) is how to minimize error during training. This is especially important in case of HMM training for medical image segmentation since incorrect results can put the life of patients at risk. In this paper, we perform extensive experimentation with HMM training for 3D brain MRI segmentation using the popular Minimum Classification Error (MCE) algorithm. We show that the behavior of such a training process is unpredictable and that no matter how long training proceeds, we cannot guarantee the decrease of the mean error. One of the promising techniques that have been proposed in the literature is Cooperative Parallel Processing (CPP) in which such training processes running in parallel on massively parallel processors broadcast the minimum global error and the corresponding parameters at regular intervals of time and resume training using these parameters. Extensive experimentation shows that CPP generally results in higher error minimization in comparison to independent training processes (that do not communicate) and that error minimization increases with the increase of the number of processors and the broadcast frequency. Nevertheless, we still cannot guarantee the decrease of the mean error during training so training has to be repeated with different broadcast frequencies to select the one that gives the best results. This led us to propose the novel CPP with Error curve Sensing (CPP-ES) technique that saves time and resources by selecting the most promising candidate broadcast frequency and avoiding higher candidate broadcast frequencies that would give worse results.