Search

Six methods for the accurate estimation of the phase-correlation maxima are discussed and evaluated in this article for one- and two-dimensional signals. The evaluation was carried out under a rigid image registration framework, where artificially generated transformations were used in order to perform a quantitative assessment of the accuracy of each method and its robustness in the presence of noise, incomplete data, or extreme transformations. Another round of tests were performed with real cases where the true transformation is unknown, and not necessarily rigid; for these tests, quantitative evaluation was achieved by means of the root mean square error of the overlapping area between the two aligned images. While most methods behaved similarly under difficult conditions, three of the methods under study displayed clear advantages under mild levels of noise, low transformation complexity, and small percentages of missing data. These methods are the local center of mass, sinc function fitting, and minimization of the POC gradient magnitude. The other tested methods included quadratic fitting, linear fitting in the frequency domain, and up-sampling; however, these methods did not perform consistently well. [ABSTRACT FROM AUTHOR]

• A-phase localization can be achieved based on abrupt EEG changes. • Localization of the onset of A-phases could help to develop a fast screening of the sleep instability. • Convolutional neural network (CNN) captures the EEG characteristics that identify the transitions and non-transitions induced by A-phases. The electroencephalogram (EEG) conveys information related to different sleep processes. One of these processes is the Cyclic Alternating Pattern (CAP), which is correlated with sleep instability. CAP is composed of A-phases, which are short recurrent modifications to the EEG fluctuations that characterize the sleep stages. A-phase annotation is performed by trained clinicians by visual EEG inspection, thus this is a weary and time-consuming task. A-phase annotation is a three step task: 1) localization, 2) delineation and 3) categorization. We propose to resolve the first step, to identify the A-phase location by training a deep convolutional neural network (CNN) based on the A-phase clinical description: an abrupt modification of the basal EEG fluctuations. Whole night EEG recordings of nine healthy subjects were used in this study. As first step, a CNN was trained and tested with the Leave-One-Out scheme in a balanced dataset of 4s EEG segments where an A-phase onset was or was not present. As a second step, the trained CNNs were used to identify A-phase onsets across the whole night recording. The results showed an accuracy performance of 93%, sensitivity of 94% and specificity of 91% for the balanced set. On the whole recording, the performance was: F-score of 58%, recall of 70% and precision of 49%. In conclusion, we present a simple fully automatic method to localize the onset of A-phases in EEG signals. It is based on the spectral characteristics of the EEG signal which define the A-phases and could be part of more complex systems. [ABSTRACT FROM AUTHOR]

Abstract: In this paper, we present a new methodology for multimodal non-rigid medical image registration. The proposed approach is based on combining a rigid registration achieved by a global optimization method, and a multimodal optical flow technique based on the conditional statistics of the joint intensity distribution (CS-JID) of the images to register. The methodology is essentially composed of two steps: first, the global deformation is approximated by using a rigid registration based on particle filtering; second, the optical flow is applied iteratively between the target and sequentially registered source image, by optimizing a new energy function that penalizes the difference between the intensities in one image with respect to the mean of the conditional intensity distribution of the other image, weighted by the conditional variance. After these steps, the non-rigid registration is made up by adding the resulting vector fields, computed by the rigid registration, and the sequential optical flow. The proposed algorithm was tested with three pairs of computed tomography (CT) and magnetic resonance (MR) images, aligned in the acquisition process, and subsequently warped with a synthetic non-rigid deformation. Preliminary results show that the methodology presents a good alternative for non-rigid multimodal registration, obtaining an average error of less than one pixel in the estimation of the deformation vector field in the majority of the cases. [Copyright &y& Elsevier]

Abstract: In this work, we present an optical flow algorithm with a highly reduced set of motion vector candidates that are obtained by the Phase-Only Correlation function. For each candidate and each pixel, we estimate the cost of assigning the candidate vector to that pixel using a sophisticated filter that is robust to edges, instead of the most commonly used box filter for block matching. The proposed filter is based on the minimization of a quadratic energy function using the Gauss-Seidel optimization method. Preliminary results with images from the Middlebury database show that the proposed method reduces the angular and end-point errors by up to 25% with respect to the classic block matching method. [Copyright &y& Elsevier]