Classification of signal-time curves from dynamic MR mammography by neural networks

The aim of this study was to test the performance of artificial neural networks for the classification of signal-time curves obtained from breast masses by dynamic MRI. Signal-time courses from 105 parenchyma, 162 malignant, and 102 benign tissue regions were examined. The latter two groups were histopathologically verified. Four neural networks corresponding to different temporal resolutions of the signal-time curves were tested. The resolution ranges from 28 measurements with a temporal spacing of 23s to just 3 measurements taken 1.8, 3, and 10 minutes after contrast medium administration. Discrimination between malignant and benign lesions is best if 28 measurement points are used (sensitivity: 84%, specificity: 81%). The use of three measurement points results in 78% sensitivity and 76% specificity. These results correspond to values obtained by human experts who visually evaluated signal-time curves without considering additional morphologic information. All examined networks yielded poor results for the subclassification of the benign lesions into fibroadenomas and benign proliferative changes. Neural networks can computationally fast distinguish between malignant and benign lesions even when only a few post-contrast measurements are made. More precise specification of the type of the benign lesion will require incorporation of additional morphological or pharmacokinetic information.     

The respiratory modulation of intracranial cerebrospinal fluid pulsation observed on dynamic echo planar images

Pressure changes in cerebrospinal fluid (CSF) that occur with respiration rhythms have been studied in animals and humans for more than 100 years. This phenomenon has been recently validated in vivo on MR images by applying spectral analysis to signal-time curves at selected regions of interest. However, selecting regions of interest requires knowledge of physiology and anatomy, and manual selection is time consuming. We postulate that CSF pulsation is passively modulated by intra-thoracic pressure that is secondary to respiration, and this pulsation can be observed as a flow-related enhancement on MR images. To investigate the spatiotemporal patterns of respiratory rhythms in human brains, we conducted a study on MR scanning of 12 healthy volunteers who performed normal-breathing and breath-holding experiments during scanning. Spectral analysis, spectroscopic images, independent component analysis and signal measurements in selected regions were applied to dynamic MR images acquired from these volunteers. Through independent component analysis, respiratory rhythms were found at the vicinity of ventricles and CSF areas in nine subjects in normal-breathing experiments. In breath-holding experiments, respiratory rhythm suppression and vessel dilation were observed in 8 and 10 subjects, respectively. Information obtained from this study further elucidates the respiratory modulation of CSF in vivo.     

Detection of areas with viable remnant tumor in postchemotherapy patients with Ewing's sarcoma by dynamic contrast-enhanced MRI using pharmacokinetic modeling

An approach is presented for monitoring the effects of neoadjuvant chemotherapy in patients with Ewing's sarcoma using dynamic contrast-enhanced perfusion magnetic resonance (MR) images. For that purpose, we modify the three-compartment pharmacokinetic permeability model introduced by Tofts et al. (Magn Reson Med 1991;17:357-67) to a two-compartment model. Perfusion MR images acquired using an intravenous injection with Gadolinium (Gd-DTPA) are analyzed with this two-compartment pharmacokinetic model as well as the with an extended pharmacokinetic model that includes the (local) arrival time t(0) of the tracer as an endogenous (estimated) parameter. For each MR section, a wash-in parameter associated with each voxel is estimated twice by fitting each of the two pharmacokinetic models to the dynamic MR signal. A comparison of the two wash-in parametric images (global versus local arrival time) with matched histologic macroslices demonstrates a good correspondence between areas with viable remnant tumor and a high wash-in rate. This can be explained by the high number and permeability of the (leaking) capillaries in viable tumor tissue. The novel pharmacokinetic model based on a local arrival time of tracer results in the best fit of the wash-in rate, the most important factor discerning viable from nonviable tumor components. However, parameter estimates obtained with this model are also more sensitive to noise in the MR signal. The novel pharmacokinetic model resulted in a sensitivity between 0.22 and 0.60 and a specificity between 0.61 and 1. The model based on a global arrival time gave sensitivities between 0.33 and 0.77 and specificities between 0.58 and 0.99. Both statistics are computed as the fraction of correctly labeled voxels (viable or nonviable tumor) within a specified ROI, which delineates the tumor. We conclude that the added value of estimating the local arrival time of tracer first manifests itself for moderate noise levels in the MR signal. The novel pharmacokinetic model should moreover be preferred when pharmacokinetic modeling is applied on the average signal intensity within a ROI, where noise has less effect on the fitted parameters.     

ECG Signal Classification Using Deep Learning Techniques Based on the PTB-XL Dataset

The analysis and processing of ECG signals are a key approach in the diagnosis of cardiovascular diseases. The main field of work in this area is classification, which is increasingly supported by machine learning-based algorithms. In this work, a deep neural network was developed for the automatic classification of primary ECG signals. The research was carried out on the data contained in a PTB-XL database. Three neural network architectures were proposed: the first based on the convolutional network, the second on SincNet, and the third on the convolutional network, but with additional entropy-based features. The dataset was divided into training, validation, and test sets in proportions of 70%, 15%, and 15%, respectively. The studies were conducted for 2, 5, and 20 classes of disease entities. The convolutional network with entropy features obtained the best classification result. The convolutional network without entropy-based features obtained a slightly less successful result, but had the highest computational efficiency, due to the significantly lower number of neurons.     

Enhancing the acquisition efficiency of fast magnetic resonance imaging via broadband encoding of signal content

Current efficient magnetic resonance imaging (MRI) methods such as parallel-imaging and k-t methods encode MR signals using a set of effective encoding functions other than the Fourier basis. This work revisits the proposition of directly manipulating the set of effective encoding functions at the radiofrequency excitation step in order to increase MRI efficiency. This approach, often termed "broadband encoding," enables the application of algebraic matrix factorization technologies to extract efficiency by representing and encoding MR signal content in a compacted form. Broadband imaging equivalents of fast multiecho, parallel and k-t MRI are developed and analyzed. The potential of these techniques to increase the time efficiency of data acquisition is experimentally verified on a commercial MRI scanner using simple spin-echo imaging. A three-dimensional gradient-echo dynamic imaging application that demonstrates the potential benefits of this approach compared to the present state of the art for certain applications is also presented.     

Two-dimensional spectroscopic imaging for pretreatment evaluation of prostate cancer: comparison with the step-section histology after radical prostatectomy

Purpose

To minimize user and vendor dependence of the spectrum processing of prostate spectra, to measure the ratio of choline (Cho) plus creatine (Cr) to citrate (Cit) in the prostate tissue of normal volunteers and cancer patients, and to compare the results with pathologic findings after radical prostatectomy.

Materials and methods

Four healthy volunteers and 13 patients with prostate cancer were measured. Measurements were performed using two-dimensional magnetic resonance spectroscopic imaging (MRSI) and endorectal coil. A standard vendor's spectrum processing approach has been modified. An original feature of this methodology was the combination of vendor-optimized and user-independent spectrum preprocessing in the scanner and user-independent quantitation in the environment of an MRUI software package. (Cho+Cr)/Cit ratio was used for the classification of prostate tissue. Results were compared with histopathology after radical prostatectomy.

Results

Eight of 13 cancer patients were classified as suspicious or very suspicious for cancer at spectroscopy, three were ambiguous for cancer and two patients were evaluated as false negative. A considerable overlap of metabolite ratios at various Gleason score was found.

Conclusion

The proposed spectrum processing has the potential to improve the accuracy and user independency of the (Cho+Cr)/Cit quantitation. This study confirmed the previous results that a considerable overlap of (Cho+Cr)/Cit ratios exists at various Gleason score levels.     

One- and Two-Phase Software Requirement Classification Using Ensemble Deep Learning

Recently, deep learning (DL) has been utilized successfully in different fields, achieving remarkable results. Thus, there is a noticeable focus on DL approaches to automate software engineering (SE) tasks such as maintenance, requirement extraction, and classification. An advanced utilization of DL is the ensemble approach, which aims to reduce error rates and learning time and improve performance. In this research, three ensemble approaches were applied: accuracy as a weight ensemble, mean ensemble, and accuracy per class as a weight ensemble with a combination of four different DL models—long short-term memory (LSTM), bidirectional long short-term memory (BiLSTM), a gated recurrent unit (GRU), and a convolutional neural network (CNN)—in order to classify the software requirement (SR) specification, the binary classification of SRs into functional requirement (FRs) or non-functional requirements (NFRs), and the multi-label classification of both FRs and NFRs into further experimental classes. The models were trained and tested on the PROMISE dataset. A one-phase classification system was developed to classify SRs directly into one of the 17 multi-classes of FRs and NFRs. In addition, a two-phase classification system was developed to classify SRs first into FRs or NFRs and to pass the output to the second phase of multi-class classification to 17 classes. The experimental results demonstrated that the proposed classification systems can lead to a competitive classification performance compared to the state-of-the-art methods. The two-phase classification system proved its robustness against the one-phase classification system, as it obtained a 95.7% accuracy in the binary classification phase and a 93.4% accuracy in the second phase of NFR and FR multi-class classification.     

Efficiency of autonomous soft nanomachines at maximum power

We consider nanosized artificial or biological machines working in steady state enforced by imposing nonequilibrium concentrations of solutes or by applying external forces, torques, or electric fields. For unicyclic and strongly coupled multicyclic machines, efficiency at maximum power is not bounded by the linear response value 1/2. For strong driving, it can even approach the thermodynamic limit 1. Quite generally, such machines fall into three different classes characterized, respectively, as "strong and efficient," "strong and inefficient," and "balanced." For weakly coupled multicyclic machines, efficiency at maximum power has lost any universality even in the linear response regime.     

Using Single-Channel Blind Deconvolution to Choose the Most Realistic Pharmacokinetic Model in Dynamic Contrast-Enhanced MR Imaging

In dynamic contrast-enhanced magnetic resonance imaging, there has been no consensus in the choice of the pharmacokinetic model. In this paper, a new approach for assessment of the most realistic model for a given tissue is presented. Non-blind and single-channel blind deconvolution algorithms were used in quantitative magnetic resonance dynamic contrast-enhanced imaging of the mouse masseter muscle to compare the realism of two different pharmacokinetic models for the tissue residue function. The first was the adiabatic approximation tissue homogeneity model (aaJW) and the second, the two-compartment exchange model (2CXM). Normals and mice treated with the substance C48/80 were studied. C48/80 increases both blood flow and contrast leakage in muscle substantially. The obtained approximation accuracy was evaluated for both pharmacokinetic models. In addition, the arterial input functions (aifs) estimated using blind deconvolution were compared to the corresponding observed aifs. The hypothesis is that the most realistic model of the tissue residue function leads to the best fits. The non-blind deconvolution did not result in any clear answer. For blind deconvolution, the aifs of the aaJW model were very similar to the corresponding observed aifs, and clearly more so than the aifs of the 2CXM model. Also, the approximation of the observed tracer time sequences was more accurate for the aaJW than the 2CXM model. The realism of different pharmacokinetic models in describing the passage of a tracer through a microvascular bed of a single tissue could be assessed using single-channel blind deconvolution.     

Application of fuzzy c-means segmentation technique for tissue differentiation in MR images of a hemorrhagic glioblastoma multiforme

The application of a raw data-based, operator-independent MR segmentation technique to differentiate boundaries of tumor from edema or hemorrhage is demonstrated. A case of a glioblastoma multiforme with gross and histopathologic correlation is presented. The MR image data set was segmented into tissue classes based on three different MR weighted image parameters (T₁-, proton density-, and T₂-weighted) using unsupervised fuzzy c-means (FCM) clustering algorithm technique for pattern recognition. A radiological examination of the MR images and correlation with fuzzy clustering segmentations was performed. Results were confirmed by gross and histopathology which, to the best of our knowledge, reports the first application of this demanding approach. Based on the results of neuropathologic correlation, the application of FCM MR image segmentation to several MR images of a glioblastoma multiforme represents a viable technique for displaying diagnostically relevant tissue contrast information used in 3D volume reconstruction. With this technique, it is possible to generate segmentation images that display clinically important neuroanatomic and neuropathologic tissue contrast information from raw MR image data.     

一种新型广义RBF神经网络在混沌时间序列预测中的研究

提出了一种新颖的广义径向基函数神经网络模型,其径向基函数(RBF)的形式由生成函数确定.然后,给出了易实现的梯度学习算法,同时为了进一步提高网络的收敛速度和网络性能,又给出了基于卡尔曼滤波的动态学习算法.为了验证网络的学习性能，采用基于卡尔曼滤波算法的新型广义RBF网络预测模型对Mackey-Glass混沌时间序列和Henon映射进行了仿真.结果表明,所提出的新型广义RBF神经网络模型能快速、精确地预测混沌时间序列,是研究复杂非线性动力系统辨识和控制的一种有效方法. 关键词： 广义径向基函数神经网络 卡尔曼滤波 梯度下降学习算法 混沌时间序列 预测     

基于线性预测倒谱的被动声纳目标特征提取技术

从声纳员的角度出发，被动声纳目标可以被看作为一个发声体，利用线性预测倒谱从声纳目标噪声中分离出目标作为发声体的冲激响应在倒谱域中的表示，提取一组识别特征，设计神经网络分类器，对三类目标进行分类。实测数据验证了基于线性预测倒谱的被动声纳目标特征提取方法是可行的。     

Permeability Coefficients from NMR q-Space Data: Models with Unevenly Spaced Semi-permeable Parallel Membranes

The NMR "q-space" experiment conducted on water provides information on the sizes of repeated structures on the micrometer-length scale in heterogeneous samples, including cell suspensions or tissues. Under some circumstances these plots display coherence peaks, and it has been implied theoretically that the position of the peaks will vary with the rate of molecular exchange across the membranes. This has been demonstrated (qualitatively) with human erythrocytes in suspension. Thus, in the quest for a quantitative approach to the interpretation of such data, we address here the "inverse problem," namely the estimate of the permeability coefficient of membranes from q-space experiments. The present work describes theoretical predictions of q-space plots from molecules diffusing in a simple system of parallel semi-permeable membranes arranged with separations that alternate between two different values; this was designed to (loosely) mimic the intra- and extracellular compartments in a suspension of cells or a tissue. The development of the theory was facilitated by symbolic computation, and the analysis of synthetic data was shown to be achievable by the use of a three-layer back-propagation artificial neural network.     

A 3D densely connected convolution neural network with connection-wise attention mechanism for Alzheimer's disease classification

PurposeAlzheimer's disease (AD) is a progressive and irreversible neurodegenerative disease. In recent years, machine learning methods have been widely used on analysis of neuroimage for quantitative evaluation and computer-aided diagnosis of AD or prediction on the conversion from mild cognitive impairment (MCI) to AD. In this study, we aimed to develop a new deep learning method to detect or predict AD in an efficient way.Materials and methodsWe proposed a densely connected convolution neural network with connection-wise attention mechanism to learn the multi-level features of brain MR images for AD classification. We used the densely connected neural network to extract multi-scale features from pre-processed images, and connection-wise attention mechanism was applied to combine connections among features from different layers to hierarchically transform the MR images into more compact high-level features. Furthermore, we extended the convolution operation to 3D to capture the spatial information of MRI. The features extracted from each 3D convolution layer were integrated with features from all preceding layers with different attention, and were finally used for classification. Our method was evaluated on the baseline MRI of 968 subjects from ADNI database to discriminate (1) AD versus healthy subjects, (2) MCI converters versus healthy subjects, and (3) MCI converters versus non-converters.ResultsThe proposed method achieved 97.35% accuracy for distinguishing AD patients from healthy control, 87.82% for MCI converters against healthy control, and 78.79% for MCI converters against non-converters. Compared with some neural networks and methods reported in recent studies, the classification performance of our proposed algorithm was among the top ranks and improved in discriminating MCI subjects who were in high risks of conversion to AD.ConclusionsDeep learning techniques provide a powerful tool to explore minute but intricate characteristics in MR images which may facilitate early diagnosis and prediction of AD.     

Are signal intensity and homogeneity useful parameters for distinguishing between benign and malignant soft tissue masses on MR images?: Objective evaluation by means of texture analysis

Objectives

To objectively identify possible differences in the signal characteristics of benign and malignant soft tissue masses (STM) on magnetic resonance (MR) images by means of texture analysis and to determine the value of these differences for computer-assisted lesion classification.

Method

Fifty-eight patients with histologically proven STM (benign, n=30; malignant, n=28) were included. STM texture was analyzed on routine T1-weighted, T2-weighted and short tau inversion recovery (STIR) images obtained with heterogeneous acquisition protocols. Fisher coefficients (F) and the probability of classification error and average correlation coefficients (POE+ACC) were calculated to identify the most discriminative texture features for separation of benign and malignant STM. F>1 indicated adequate discriminative power of texture features. Based on the texture features, computer-assisted classification of the STM by means of k-nearest-neighbor (k-NN) and artificial neural network (ANN) classification was performed, and accuracy, sensitivity and specificity were calculated.

Results

Discriminative power was only adequate for two texture features, derived from the gray-level histogram of the STIR images (first and 10th gray-level percentiles). Accordingly, the best results of STM classification were achieved using texture information from STIR images, with an accuracy of 75.0% (sensitivity, 71.4%; specificity, 78.3%) for the k-NN classifier, and an accuracy of 90.5% (sensitivity, 91.1%; specificity, 90.0%) for the ANN classifier.

Conclusion

Texture analysis revealed only small differences in the signal characteristics of benign and malignant STM on routine MR images. Computer-assisted pattern recognition algorithms may aid in the characterization of STM, but more data is necessary to confirm their clinical value.     

Categorizing Normal and Pathological Voices: Automated and Perceptual Categorization

Objectives

The aims of the present study were to evaluate the accuracy of an elaborated automated voice categorization system that classified voice signal samples into healthy and pathological classes and to compare it with classification accuracy that was attained by human experts.

Material and Methods

We investigated the effectiveness of 10 different feature sets in the classification of voice recordings of the sustained phonation of the vowel sound /a/ into the healthy and two pathological voice classes, and proposed a new approach to building a sequential committee of support vector machines (SVMs) for the classification. By applying “genetic search” (a search technique used to find solutions to optimization problems), we determined the optimal values of hyper-parameters of the committee and the feature sets that provided the best performance. Four experienced clinical voice specialists who evaluated the same voice recordings served as experts. The “gold standard” for classification was clinically and histologically proven diagnosis.

Results

A considerable improvement in the classification accuracy was obtained from the committee when compared with the single feature type-based classifiers. In the experimental investigations that were performed using 444 voice recordings coming from 148 subjects, three recordings from each subject, we obtained the correct classification rate (CCR) of over 92% when classifying into the healthy-pathological voice classes, and over 90% when classifying into three classes (healthy voice and two nodular or diffuse lesion voice classes). The CCR obtained from human experts was about 74% and 60%, respectively.

Conclusion

When operating under the same experimental conditions, the automated voice discrimination technique based on sequential committee of SVM was considerably more effective than the human experts.     

MRI texture analysis on texture test objects, normal brain and intracranial tumors

Texture analysis was performed in three different MRI units on T1 and T2-weighted MR images from 10 healthy volunteers and 63 patients with histologically confirmed intracranial tumors. The goal of this study was a multicenter evaluation of the usefulness of this quantitative approach for the characterization of healthy and pathologic human brain tissues (white matter, gray matter, cerebrospinal fluid, tumors and edema). Each selected brain region of interest was characterized with both its mean gray level values and several texture parameters. Multivariate statistical analyses were then applied in order to discriminate each brain tissue type represented by its own set of texture parameters. Texture analysis was previously performed on test objects to evaluate the method dependence on acquisition parameters and consequently the interest of a multicenter evaluation. Even obtained on different sites with their own acquisition routine protocol, MR brain images contain textural features that can reveal discriminant factors for tissue classification and image segmentation. It can also offer additional information in case of undetermined diagnosis or to develop a more accurate tumor grading.     

Segmentation of acoustic images by neural network processing

A segmentation method for biomedical acoustic images is reported which efficiently classifies the groups of similar image elements (pixels) and separates them into particular characteristic regions. As the input data, the method uses the pixel intensities of the source image. The classification is performed by learning vector quantization neural networks, which separate the main classes (structures, tissues, artifacts, etc.) present in the image. Because this type of neural network implies that the number of the classes is known and that the network should be trained by instruction, an expert must participate in the process of generating the input data. Results obtained by processing test acoustic (ultrasonic) images demonstrate that the method is capable of effectively solving sonography classification problems. The accuracy of the method is estimated by comparison with the segmentation performed manually.     

A physical approach to the automated classification of clinical percussion sounds

Chest percussion is a traditional technique used for the physical examination of pulmonary injuries and diseases. It is a method of tapping body parts with fingers or small instruments to evaluate the size, consistency, borders, and presence of fluid/air in the lungs and abdomen. Percussion has been successfully used for the diagnosis of such potentially lethal conditions as traumatic and tension pneumothorax. This technique, however, has certain shortcomings, including limitations of the human ear and the subjectivity of the administrator, that lead to overall low sensitivity. Automation of the method by using a standardized percussion source and computerized classification of digitized signals would remove the subjective factor and other limitations of the technique. It would also enable rapid on-site diagnostics of pulmonary traumas when thorough clinical examination is impossible. This paper lays the groundwork for an objective signal classification approach based on a general physical model of a damped harmonic oscillator. Using this concept, critical parameters that effectively subdivide percussion signals into three main groups, historically known as "tympanic," "resonant," and "dull," are identified, opening the possibility for automated diagnostics of air/liquid inclusions in the thorax and abdomen. The key role of damping in forming the character of the percussion signal is investigated using a 3D thorax phantom. The contribution of the abdominal component into the complex multimode spectrum of chest percussion signals is demonstrated.     

从二维视图识别三维目标的多网络融合方法

提出了一种从二维视图识别三维目标的多网络融合方法,基于单个网络分类的置信度概念,有效地结合多个网络的输出结果作出最终分类判决,应用三个多层前向网络（隐层神经元数,初始权值等取不同值）,设计了基于分类确认度的多网络融合结构,对四类车辆目标进行的识别实验表明,所提出的多网络融合方法明显优于单个网络的识别性能。