Axel Wismüller

Cluster Analysis of Biomedical Image Time-Series

In this paper, we present neural network clustering by deterministic annealing as a powerful strategy for self-organized segmentation of biomedical image time-series data identifying groups of pixels sharing common properties of local signal dynamics. After introducing the theoretical concept of minimal free energy vector quantization and related clustering techniques, we discuss its potential to serve as a multi-purpose computer vision strategy to image time-series analysis and visualization for many fields of medicine ranging from biomedical basic research to clinical assessment of patient data. In particular, we present applications to (i) functional MRI data analysis for human brain mapping, (ii) dynamic contrast-enhanced perfusion MRI for the diagnosis of cerebrovascular disease, and (iii) magnetic resonance mammography for the analysis of suspicious lesions in patients with breast cancer. This wide scope of completely different medical applications illustrates the flexibility and conceptual power of neural network vector quantization in this context. Although there are obvious methodological similarities, each application requires specific careful consideration w.r.t. data preprocessing, postprocessing and interpretation. This challenge can only be managed by close interdisciplinary cooperation of medical doctors, engineers, and computer scientists. Hence, this field of research can serve as an example for lively cross-fertilization between computer vision and related research.

Monthly intravenous methylprednisolone in relapsing-remitting multiple sclerosis - reduction of enhancing lesions, T2 lesion volume and plasma prolactin concentrations

BackgroundIntravenous methylprednisolone (IV-MP) is an established treatment for multiple sclerosis (MS) relapses, accompanied by rapid, though transient reduction of gadolinium enhancing (Gd+) lesions on brain MRI. Intermittent IV-MP, alone or with immunomodulators, has been suggested but insufficiently studied as a strategy to prevent relapses.MethodsIn an open, single-cross-over study, nine patients with relapsing-remitting MS (RR-MS) underwent cranial Gd-MRI once monthly for twelve months. From month six on, they received a single i.v.-infusion of 500 mg methylprednisolone (and oral tapering for three days) after the MRI. Primary outcome measure was the mean number of Gd+ lesions during treatment vs. baseline periods; T2 lesion volume and monthly plasma concentrations of cortisol, ACTH and prolactin were secondary outcome measures. Safety was assessed clinically, by routine laboratory and bone mineral density measurements. Soluble immune parameters (sTNF-RI, sTNF-RII, IL1-ra and sVCAM-1) and neuroendocrine tests (ACTH test, combined dexamethasone/CRH test) were additionally analyzed.ResultsComparing treatment to baseline periods, the number of Gd+ lesions/scan was reduced in eight of the nine patients, by a median of 43.8% (p = 0.013, Wilcoxon). In comparison, a pooled dataset of 83 untreated RR-MS patients from several studies, selected by the same clinical and MRI criteria, showed a non-significant decrease by a median of 14% (p = 0.32). T2 lesion volume decreased by 21% during treatment (p = 0.001). Monthly plasma prolactin showed a parallel decline (p = 0.027), with significant cross-correlation with the number of Gd+ lesions. Other hormones and immune system variables were unchanged, as were ACTH test and dexamethasone-CRH test. Treatment was well tolerated; routine laboratory and bone mineral density were unchanged.ConclusionMonthly IV-MP reduces inflammatory activity and T2 lesion volume in RR-MS.

Adaptive local dissimilarity measures for discriminative dimension reduction of labeled data

Due to the tremendous increase of electronic information with respect to the size of data sets as well as their dimension, dimension reduction and visualization of high-dimensional data has become one of the key problems of data mining. Since embedding in lower dimensions necessarily includes a loss of information, methods to explicitly control the information kept by a specific dimension reduction technique are highly desirable. The incorporation of supervised class information constitutes an important specific case. The aim is to preserve and potentially enhance the discrimination of classes in lower dimensions. In this contribution we use an extension of prototype-based local distance learning, which results in a nonlinear discriminative dissimilarity measure for a given labeled data manifold. The learned local distance measure can be used as basis for other unsupervised dimension reduction techniques, which take into account neighborhood information. We show the combination of different dimension reduction techniques with a discriminative similarity measure learned by an extension of learning vector quantization (LVQ) and their behavior with different parameter settings. The methods are introduced and discussed in terms of artificial and real world data sets.

Cluster analysis of dynamic cerebral contrast-enhanced perfusion MRI time-series

We performed neural network clustering on dynamic contrast-enhanced perfusion magnetic resonance imaging time-series in patients with and without stroke. Minimal-free-energy vector quantization, self-organizing maps, and fuzzy c-means clustering enabled self-organized data-driven segmentation with respect to fine-grained differences of signal amplitude and dynamics, thus identifying asymmetries and local abnormalities of brain perfusion. We conclude that clustering is a useful extension to conventional perfusion parameter maps.

Classification of Small Lesions in Breast MRI: Evaluating The Role of Dynamically Extracted Texture Features Through Feature Selection

Dynamic texture quantification, i.e., extracting texture features from the lesion enhancement pattern in all available post-contrast images, has not been evaluated in terms of its ability to classify small lesions. This study investigates the classification performance achieved with texture features extracted from all five post-contrast images of lesions (mean lesion diameter of 1.1 cm) annotated in dynamic breast magnetic resonance imaging exams. Sixty lesions are characterized dynamically using Haralick texture features. The texture features are then used in a classification task with support vector regression and a fuzzy k-nearest neighbor classifier; free parameters of these classifiers are optimized using random sub-sampling cross-validation. Classifier performance is determined through receiver-operator characteristic (ROC) analysis, specifically through computation of the area under the ROC curve (AUC). Mutual information is used to evaluate the contribution of texture features extracted from different post-contrast stages to classifier performance. Significant improvements (p < 0.05) are observed for six of the thirteen texture features when the lesion enhancement pattern is quantified using the proposed approach of dynamic texture quantification. The highest AUC value observed (0.82) is achieved with texture features responsible for capturing aspects of lesion heterogeneity. Mutual information analysis reveals that texture features extracted from the third and fourth post-contrast images contributed most to the observed improvement in classifier performance. These results show that the performance of automated character classification with small lesions can be significantly improved through dynamic texture quantification of the lesion enhancement pattern.

Tumor feature visualization with unsupervised learning

Dynamic contrast enhanced magnetic resonance imaging (DCE MRI) is applied for diagnosis and therapy control of breast cancer. The malignancy of a lesion is expressed in the average signal kinetics of selected regions of interest (ROI) representing the lesion. The technique is reported to characterize malignant tumors with high sensitivity and highly variable specificity. Computer-based diagnosis (CAD) systems have been proposed to analyze and classify signal time curve data, extracted from hand selected ROI in the DCE MRI data. In this paper, we apply the self-organizing map (SOM) to a set of time curve feature vectors of single voxels from seven benign lesions and seven malignant tumors. Applying the SOM we are able to project the time curve values of each voxel on a two-dimensional map. The results show, that the SOM is able to visualize the hidden two-dimensional structure of the six-dimensional signal space. Using the trained SOM, we are able to identify voxels with benign or malignant signal characteristics and to visualize lesion cross-sections with pseudo-colors. A comparison with the established three time points method shows that the SOM has clear potential for deriving visualization parameters in DCE MRI analysis.

Classification of Small Contrast Enhancing Breast Lesions in Dynamic Magnetic Resonance Imaging Using a Combination of Morphological Criteria and Dynamic Analysis Based on Unsupervised Vector-Quantization

Purpose:To evaluate the diagnostic value of breast magnetic resonance imaging (MRI) in small focal lesions using dynamic analysis based on unsupervised vector quantization in combination with a score for morphologic criteria. Materials and Methods:We examined 85 mammographically indetermintate lesions (BIRADS 3–4; 47 malignant, mean lesion size 1.2 cm; 38 benign, mean lesion size 1.1 cm). MRI was performed with a dynamic T1-weighted gradient echo sequence (1 precontrast and 5 postcontrast series). Lesions with an initial contrast enhancement ≥50% were selected with semiautomatic segmentation. For conventional dynamic analysis, we calculated the mean initial signal increase and postinitial course of all voxels included in a lesion. Secondly, all voxels within the lesions were assigned to 4 clusters using minimal-free-energy vector quantization. Dynamic and morphologic criteria were summarized in a diagnostic score and evaluated by receiver operating characteristic analysis. Results:In the present collection of small lesions, morphologic criteria [area under the curve (AUC) = 0.610] were inferior to dynamic criteria in the detection of breast cancer. Dynamic analysis with vector quantization (AUC = 0.760) presented slightly better results compared with standard dynamic analysis (AUC = 0.693). There was no benefit for combined morphologic and dynamic analysis. Conclusion:In small MR-mammographic lesions, dynamic analysis with vector quantization alone tends to result in a higher diagnostic accuracy compared with combined morphologic and dynamic analysis.

Neighbor embedding XOM for dimension reduction and visualization

We present an extension of the Exploratory Observation Machine (XOM) for structure-preserving dimensionality reduction. Based on minimizing the Kullback-Leibler divergence of neighborhood functions in data and image spaces, this Neighbor Embedding XOM (NE-XOM) creates a link between fast sequential online learning known from topology-preserving mappings and principled direct divergence optimization approaches. We quantitatively evaluate our method on real-world data using multiple embedding quality measures. In this comparison, NE-XOM performs as a competitive trade-off between high embedding quality and low computational expense, which motivates its further use in real-world settings throughout science and engineering.

The deformable feature map - a novel neurocomputing algorithm for adaptive plasticity in pattern analysis

Abstract In this paper, we present an algorithm that provides adaptive plasticity in function approximation problems: the deformable (feature) map (DM) algorithm. The DM approach reduces a class of similar function approximation problems to the explicit supervised one-shot training of a single data set. This is followed by a subsequent, appropriate similarity transformation which is based on a self-organized deformation of the underlying multi-dimensional probability distributions. After discussing the theory of the DM algorithm, we use computer simulations to visualize its effects on low-dimensional toy examples. Finally, we present results of its application to the real-world problem of automatic nonlinear multispectral image registration, employing magnetic resonance data sets of the human brain.

Local exponential stability of competitive neural networks with different time scales

Abstract This contribution presents a new method of analyzing the dynamics of a biological relevant neural network with different time scales based on the theory of flow invariance. We are able to show that the resulting stability conditions are less restrictive and more general than with K -monotone theory or singular perturbation theory. The theoretical results are further substantiated by simulation results conducted for analysis and design of these neural networks.