Ultrasonic tissue characterization has become an area of intensive research. This procedure generally relies on the analysis of the unprocessed echo signal. Because the ultrasound echo is degraded by the non-ideal system point spread function, a deconvolution step could be employed to provide an estimate of the tissue response that could then be exploited for a more accurate characterization. In medical ultrasound, deconvolution is commonly used to increase diagnostic reliability of ultrasound images by improving their contrast and resolution. Most successful algorithms address deconvolution in a maximum a posteriori estimation framework; this typically leads to the solution of l(2)-norm or (1)-norm constrained optimization problems, depending on the choice of the prior distribution. Although these techniques are sufficient to obtain relevant image visual quality improvements, the obtained reflectivity estimates are, however, not appropriate for classification purposes. In this context, we introduce in this paper a maximum a posteriori deconvolution framework expressly derived to improve tissue characterization. The algorithm overcomes limitations associated with standard techniques by using a nonstandard prior model for the tissue response. We present an evaluation of the algorithm performance using both computer simulations and tissue-mimicking phantoms. These studies reveal increased accuracy in the characterization of media with different properties. A comparison with state-of-the-art Wiener and l(1)-norm deconvolution techniques attests to the superiority of the proposed algorithm.

Computer-aided detection (CAD) schemes are decision making support tools, useful to overcome limitations of problematic clinical procedures. Trans-rectal ultrasound image based CAD would be extremely important to support prostate cancer diagnosis. An effective approach to realize a CAD scheme for this purpose is described in this work, employing a multi-feature kernel classification model based on generalized discriminant analysis. The mutual information of feature value and tissue pathological state is used to select features essential for tissue characterization. System-dependent effects are reduced through predictive deconvolution of the acquired radio-frequency signals. A clinical study, performed on ground truth images from biopsy findings, provides a comparison of the classification model applied before and after deconvolution, showing in the latter case a significant gain in accuracy and area under the receiver operating characteristic curve.

Abstract-The problem of estimating continuous-domain autoregressive moving-average processes from sampled data is considered. The proposed approach incorporates the sampling process into the problem formulation while introducing exponential models for both the continuous and the sampled processes. We derive an exact evaluation of the discrete-domain power-spectrum using exponential B-splines and further suggest an estimation approach that is based on digitally filtering the available data. The proposed functional, which is related to Whittle's likelihood function, exhibits several local minima that originate from aliasing. The global minimum, however, corresponds to a maximum-likelihood estimator, regardless of the sampling step. Experimental results indicate that the proposed approach closely follows the Cramér-Rao bound for various aliasing configurations.

Deep neural networks have been shown to be vulnerable to adversarial examples: very small perturbations of the input having a dramatic impact on the predictions. A wealth of adversarial attacks and distance metrics to quantify the similarity between natural and adversarial images have been proposed, recently enlarging the scope of adversarial examples with geometric transformations beyond pixel-wise attacks. In this context, we investigate the robustness to adversarial attacks of new Convolutional Neural Network architectures providing equivariance to rotations. We found that rotation-equivariant networks are significantly less vulnerable to geometric-based attacks than regular networks on the MNIST, CIFAR-10, and ImageNet datasets.

No abstract