Feasibility of infrared spectroscopy with pattern recognition techniques to identify a subpopulation of mares at risk of producing foals diagnosed with failure of transfer of passive immunity.

OBJECTIVE: To assess the feasibility of a serum-based test using infrared spectroscopy to identify a subpopulation of mares at risk of producing foals susceptible to failure of passive transfer of immunity (FPT) because of mare-associated factors. MATERIALS AND METHODS: Serum was collected from post-parturient mares (n = 126) and their foals at 24-72 h of age. A radial immunodiffusion IgG test was used to determine each foal's serum IgG concentration. Infrared absorbance spectra of dam sera were collected in the wave number range of 400-4000 cm(-1). Following data preprocessing, pattern recognition techniques were used to identify spectroscopic information capable of distinguishing between mares with FPT foals and those with normal foals. The sensitivity and specificity of infrared spectroscopy to detect risk-positive mares were calculated. RESULTS: Five wave number regions were identified as optimal for distinguishing between the two groups of mares: 740.9-785.2 cm(-1), 796.8-816.0 cm(-1), 970.4-993.5 cm(-1), 1371.6-1406.3 cm(-1) and 1632.0-1659.0 cm(-1). Based upon the infrared spectroscopic information within these discriminatory subregions, the spectra provided the risk status of the mares with a classification success rate of 81.0%. The sensitivity of the classification system was 85.7% and specificity was 80.0%. CONCLUSION: This preliminary study demonstrates that infrared spectra of dam serum have the potential to provide the basis for a new periparturient screening method for a subpopulation of mares at risk of having a foal susceptible to FPT. Further development may provide an economic and rapid technique for the pre-parturient assessment of mares.

Class proximity measures--dissimilarity-based classification and display of high-dimensional data.

For two-class problems, we introduce and construct mappings of high-dimensional instances into dissimilarity (distance)-based Class-Proximity Planes. The Class Proximity Projections are extensions of our earlier relative distance plane mapping, and thus provide a more general and unified approach to the simultaneous classification and visualization of many-feature datasets. The mappings display all L-dimensional instances in two-dimensional coordinate systems, whose two axes represent the two distances of the instances to various pre-defined proximity measures of the two classes. The Class Proximity mappings provide a variety of different perspectives of the dataset to be classified and visualized. We report and compare the classification and visualization results obtained with various Class Proximity Projections and their combinations on four datasets from the UCI data base, as well as on a particular high-dimensional biomedical dataset.

Creating robust, reliable, clinically relevant classifiers from spectroscopic data.

I describe in detail the intimately connected feature extraction and classifier development stages of the data-driven Statistical Classification Strategy (SCS) and compare them with current practice used in MR spectroscopy. We initially created the SCS for the analysis of MR and IR spectra of biofluids and tissues, and subsequently extended it to analyze biomedical data in general. I focus on explaining how to extract discriminatory spectral features and create robust classifiers that can reliably discriminate diseases and disease states. I discuss our approach to identifying features that retain spectral identity and provisionally relate these features, averaged subregions of the spectra, to specific chemical entities ("metabolites"). Particular emphasis is placed on describing the steps required to help create classifiers whose accuracy doesn't deteriorate significantly when presented with new, unknown samples. A simple but powerful extension of the discovered features to detect metabolite-metabolite (feature-feature) interactions is also sketched. I contrast the advantages and disadvantages of using either spectral signatures or explicit metabolite concentrations derived from the spectra as sets of discriminatory features. At present, no clear-cut preference is obvious and more objective comparisons will be needed. Finally, I argue that clinical requirements and exigencies strongly suggest adopting a two-phase approach to diagnosis/prognosis. In the first phase the emphasis ought to be on providing as accurate a diagnosis as possible, without any attempt to identify "biomarkers." That should be the goal of the second, research phase, with a view of providing prognosis on disease progression.

Direct classification of high-dimensional data in low-dimensional projected feature spaces--comparison of several classification methodologies.

Previously, we introduced a distance (similarity)-based mapping for the visualization of high-dimensional patterns and their relative relationships. The mapping preserves exactly the original distances from all points to any two reference patterns in a special two-dimensional coordinate system, the relative distance plane (RDP). We extend the RDP mapping's applicability from visualization to classification. Several of the classifiers use the RDP directly. These include the standard linear discriminant analysis (LDA), nearest neighbor classifiers, and a transvariation probabilities-based classification method that is natural in the RDP. Several reference directions can also be combined to create new coordinate systems in which arbitrary classifiers can be developed. We obtain increased confidence in the classification results by cycling through all possible reference pairs and computing a misclassification-based weighted accuracy. The classification results on several high-dimensional biomedical datasets are compared.

Mapping high-dimensional data onto a relative distance plane--an exact method for visualizing and characterizing high-dimensional patterns.

We introduce a distance (similarity)-based mapping for the visualization of high-dimensional patterns and their relative relationships. The mapping preserves exactly the original distances between points with respect to any two reference patterns in a special two-dimensional coordinate system, the relative distance plane (RDP). As only a single calculation of a distance matrix is required, this method is computationally efficient, an essential requirement for any exploratory data analysis. The data visualization afforded by this representation permits a rapid assessment of class pattern distributions. In particular, we can determine with a simple statistical test whether both training and validation sets of a 2-class, high-dimensional dataset derive from the same class distributions. We can explore any dataset in detail by identifying the subset of reference pairs whose members belong to different classes, cycling through this subset, and for each pair, mapping the remaining patterns. These multiple viewpoints facilitate the identification and confirmation of outliers. We demonstrate the effectiveness of this method on several complex biomedical datasets. Because of its efficiency, effectiveness, and versatility, one may use the RDP representation as an initial, data mining exploration that precedes classification by some classifier. Once final enhancements to the RDP mapping software are completed, we plan to make it freely available to researchers.

Class prediction and discovery using gene microarray and proteomics mass spectroscopy data: curses, caveats, cautions.

MOTIVATION: Two practical realities constrain the analysis of microarray data, mass spectra from proteomics, and biomedical infrared or magnetic resonance spectra. One is the 'curse of dimensionality': the number of features characterizing these data is in the thousands or tens of thousands. The other is the 'curse of dataset sparsity': the number of samples is limited. The consequences of these two curses are far-reaching when such data are used to classify the presence or absence of disease. RESULTS: Using very simple classifiers, we show for several publicly available microarray and proteomics datasets how these curses influence classification outcomes. In particular, even if the sample per feature ratio is increased to the recommended 5-10 by feature extraction/reduction methods, dataset sparsity can render any classification result statistically suspect. In addition, several 'optimal' feature sets are typically identifiable for sparse datasets, all producing perfect classification results, both for the training and independent validation sets. This non-uniqueness leads to interpretational difficulties and casts doubt on the biological relevance of any of these 'optimal' feature sets. We suggest an approach to assess the relative quality of apparently equally good classifiers.

A novel, direct spatio-temporal approach for analyzing fMRI experiments.

We introduce a novel approach to couple temporal similarity with spatial neighborhood information. This is achieved by concatenating the K nearest, spatially contiguous neighbors of a pixel time-course (TC) of T time-instances. This produces a new TC of (K+1)T time instances. Depending on how "nearest" is defined, we have various options. Strictly spatial nearness means augmenting a given TC by its K nearest neighbors in some canonical spatial order. A more powerful and flexible option is to order the TCs to be concatenated according to their temporal similarity to the central voxel TC. For this study, we have chosen Pearson's cross-correlation coefficient as the measure of similarity. For more than a single neighbor, two concatenation options are possible. The direct ordering option requires that the TCs to be concatenated be spatially contiguous to the central pixel. The more flexible indirect option merely demands that one of a chain of temporally similar TCs be spatially connected to the central pixel. We also apply the temporal similarity criterion to the more conventional spatial (median) filtering, and show that it gives superior result to a strict spatial filtering. The method is tested and verified on a null fMRI dataset onto which we superposed two types of "activations" with known temporal behavior and spatial location. It is also applied to a real dataset containing visual activation. We also propose a strategy, based on the flexibility of the method, to determine a consensus, "core" set of activations.

Exploring regions of interest with cluster analysis (EROICA) using a spectral peak statistic for selecting and testing the significance of fMRI activation time-series.

Much relevant information about activations and artifacts in a functional magnetic resonance imaging (fMRI) dataset can be obtained from an exploratory cluster analysis. In contrast to testing the significance of the measured experimental effect for a given model, unsupervised pattern recognition techniques, such as fuzzy clustering, often find unexpected behavior in addition to expected activations, allowing the exploitation of this element of surprise. The many artifact clusters often discovered might aid the experimenter in deciding whether the dataset is usable, whether some additional preprocessing step is required, or whether the one used has introduced spurious effects. However, clustering alone does not complete the analysis because the membership values that are generated are not indicative of the level of statistical significance with respect to the cluster activation patterns (centroids). This is of particular importance for fMRI datasets for which most time-series are "noise", with no activation patterns. We propose that an initial partition step should precede the clustering step. Only time-series that meet a certain statistical criterion (using a scaled version of Fisher's g-order statistic) are selected for clustering; this typically represents <5% of the whole brain region. The purpose of clustering is to generate a set of cluster centers that are the possible activation patterns; these are used in forming a linear model of all the time-series. The model parameter is tested for significance in both the time and frequency domains. We present a novel method of conducting these tests, which limits the number of false positives. We call the three-step process of initial partition, clustering and the two-domain significance test as exploring regions of interest with cluster analysis (EROICA).

Diagnosis and prognosis of breast cancer by magnetic resonance spectroscopy of fine-needle aspirates analysed using a statistical classification strategy.

BACKGROUND: The aim was to develop robust classifiers to analyse magnetic resonance spectroscopy (MRS) data of fine-needle aspirates taken from breast tumours. The resulting data could provide computerized, classification-based diagnosis and prognostic indicators. METHODS: Fine-needle aspirate biopsies obtained at the time of surgery for both benign and malignant breast diseases were analysed by one-dimensional proton MRS at 8.5 Tesla. Diagnostic correlation was performed between the spectra and standard pathology reports, including the presence of vascular invasion by the primary cancer and involvement of the excised axillary lymph nodes. RESULTS: Malignant tissue was distinguished from benign lesions with an overall accuracy of 93 per cent. From the same spectra, lymph node involvement was predicted with an overall accuracy of 95 per cent, and tumour vascular invasion with an overall accuracy of 94 per cent. CONCLUSION: The pathology, nodal involvement and tumour vascular invasion were predicted by computerized statistical classification of the proton MRS spectrum from a fine-needle aspirate biopsy taken from the primary breast lesion.

Magnetic resonance spectroscopy of the malignant prostate gland after radiotherapy: a histopathologic study of diagnostic validity.

PURPOSE: Accurate spatial representation of tumor clearance after conformal radiotherapy is an endpoint of clinical importance. Magnetic resonance spectroscopy (MRS) can diagnose malignancy in the untreated prostate gland through measurements of cellular metabolites. In this study we sought to describe spectral metabolic changes in prostatic tissue after radiotherapy and validate a multivariate analytic strategy (based on MRS) that could identify viable tumor. METHODS AND MATERIALS: Transrectal ultrasound-guided prostate biopsies from 35 patients were obtained 18-36 months after external beam radiotherapy. One hundred sixteen tissue specimens were subjected to 1H MRS, submitted to histopathology, and analyzed for correlation with a multivariate strategy specifically developed for biomedical spectra. RESULTS: The sensitivity and specificity of MRS in identifying a malignant biopsy were 88.9% and 92% respectively, with an overall classification accuracy of 91.4%. The diagnostic spectral regions identified by our algorithm included those due to choline, creatine, glutamine, and lipid. Citrate, an important discriminating resonance in the untreated prostate gland, was invisible in all spectra, regardless of histology. CONCLUSIONS: Although the spectral features of prostate tissue markedly change after radiotherapy, MRS combined with multivariate methods of analysis can accurately identify histologically malignant biopsies. MRS shows promise as a modality that could integrate three-dimensional measures of tumor response.

The use of 1H magnetic resonance spectroscopy in inflammatory bowel diseases: distinguishing ulcerative colitis from Crohn's disease.

OBJECTIVES: The distinction between the two major forms of inflammatory bowel diseases (IBD), i.e., ulcerative colitis (UC) and Crohn's disease is sometimes difficult and may lead to a diagnosis of indeterminate colitis. We have used 1H magnetic resonance spectroscopy (MRS) combined with multivariate methods of spectral data analysis to differentiate UC from Crohn's disease and to evaluate normal-appearing mucosa in IBD. METHODS: Colon mucosal biopsies (45 UC and 31 Crohn's disease) were submitted to 1H MRS, and multivariate analysis was applied to distinguish the two diseases. A second study was performed to test endoscopically and histologically normal biopsies from IBD patients. A classifier was developed by training on 101 spectra (76 inflamed IBD tissues and 25 normal control tissues). The spectra of 38 biopsies obtained from endoscopically and histologically normal areas of the colons of patients with IBD were put into the validation test set. RESULTS: The classification accuracy between UC and Crohn's disease was 98.6%, with only one case of Crohn's disease and no cases of UC misclassified. The diagnostic spectral regions identified by our algorithm included those for taurine, lysine, and lipid. In the second study, the classification accuracy between normal controls and IBD was 97.9%. Only 47.4% of the endoscopically and histologically normal IBD tissue spectra were classified as true normals; 34.2% showed "abnormal" magnetic resonance spectral profiles, and the remaining 18.4% could not be classified unambiguously. CONCLUSIONS: There is a strong potential for MRS to be used in the accurate diagnosis of indeterminate colitis; it may also be sensitive in detecting preclinical inflammatory changes in the colon.

EvIdent: a functional magnetic resonance image analysis system.

EvIdent (EVent IDENTification) is a user-friendly, algorithm-rich, exploratory data analysis software for quickly detecting, investigating, and visualizing novel events in a set of images as they evolve in time and/or frequency. For instance, in a series of functional magnetic resonance neuroimages, novelty may manifest itself as neural activations in a time course. The core of the system is an enhanced variant of the fuzzy c-means clustering algorithm. Fuzzy clustering obviates the need for models of the underlying requisite biological function, models that are often statistically suspect.

A wavelet-based method for improving signal-to-noise ratio and contrast in MR images.

Magnetic resonance (MR) images acquired with fast measurement often display poor signal-to-noise ratio (SNR) and contrast. With the advent of high temporal resolution imaging, there is a growing need to remove these noise artifacts. The noise in magnitude MR images is signal-dependent (Rician), whereas most de-noising algorithms assume additive Gaussian (white) noise. However, the Rician distribution only looks Gaussian at high SNR. Some recent work by Nowak employs a wavelet-based method for de-noising the square magnitude images, and explicitly takes into account the Rician nature of the noise distribution. In this article, we apply a wavelet de-noising algorithm directly to the complex image obtained as the Fourier transform of the raw k-space two-channel (real and imaginary) data. By retaining the complex image, we are able to de-noise not only magnitude images but also phase images. A multiscale (complex) wavelet-domain Wiener-type filter is derived. The algorithm preserves edges better when the Haar wavelet rather than smoother wavelets, such as those of Daubechies, are used. The algorithm was tested on a simulated image to which various levels of noise were added, on several EPI image sequences, each of different SNR, and on a pair of low SNR MR micro-images acquired using gradient echo and spin echo sequences. For the simulated data, the original image could be well recovered even for high values of noise (SNR approximately 0 dB), suggesting that the present algorithm may provide better recovery of the contrast than Nowak's method. The mean-square error, bias, and variance are computed for the simulated images. Over a range of amounts of added noise, the present method is shown to give smaller bias than when using a soft threshold, and smaller variance than a hard threshold; in general, it provides a better bias-variance balance than either hard or soft threshold methods. For the EPI (MR) images, contrast improvements of up to 8% (for SNR = 33 dB) were found. In general, the improvement in contrast was greater the lower the original SNR, for example, up to 50% contrast improvement for SNR of about 20 dB in micro-imaging. Applications of the algorithm to the segmentation of medical images, to micro-imaging and angiography (where the correct preservation of phase is important for flow encoding to be possible), as well as to de-noising time series of functional MR images, are discussed.

Wavelet domain de-noising of time-courses in MR image sequences.

Magnetic resonance images acquired with high temporal resolution often exhibit large noise artifacts, which arise from physiological sources as well as from the acquisition hardware. These artifacts can be detrimental to the quality and interpretation of the time-course data in functional MRI studies. A class of wavelet-domain de-noising algorithms estimates the underlying, noise-free signal by thresholding (or 'shrinking') the wavelet coefficients, assuming the underlying temporal noise of each pixel is uncorrelated and Gaussian. A Wiener-type shrinkage algorithm is developed in this paper, for de-noising either complex- or magnitude-valued image data sequences. Using the de-correlation properties of the wavelet transform, as elucidated by Johnstone and Silverman, the assumption of i.i.d. Gaussian noise can be abandoned, opening up the possibility of removing colored noise. Both wavelet- and wavelet-packet based algorithms are developed, and the Wiener method is compared to the traditional Hard and Soft wavelet thresholding methods of Donoho and Johnstone. The methods are applied to two types of data sets. In the first, an artificial set of complex-valued images was constructed, in which each pixel has a simulated bimodal time-course. Gaussian noise was added to each of the real and imaginary channels, and the noise removed from the complex image sequence as well as the magnitude image sequence (where the noise is Rician). The bias and variance between the original and restored paradigms was estimated for each method. It was found that the Wiener method gives better balance in bias and variance than either Hard or Soft methods. Furthermore, de-noising magnitude data provides comparable accuracy of the restored images to that obtained from de-noising complex data. In the second data set, an actual in vivo complex image sequence containing unknown physiological and instrumental noise was used. The same bimodal paradigm as in the first data set was added to pixels in a small localized region of interest. For the paradigm investigated here, the smooth Daubechies wavelets provide better de-noising characteristics than the discontinuous Haar wavelets. Also, it was found that wavelet packet de-noising offers no significant improvement over the computationally more efficient wavelet de-noising methods. For the in vivo data, it is desirable that the groups of "activated" time-courses are homogeneous. It was found that the internal homogeneity of the group of time-courses increases when de-noising is applied. This suggests using de-noising as a pre-processing tool for both exploratory and inferential data analysis methods in fMRI.

Disease pattern recognition in infrared spectra of human sera with diabetes mellitus as an example.

To benefit from the full information content of the mid-IR spectra of human sera, we directly related the overall shape of the spectra to the donors' disease states. For this approach of disease pattern recognition we applied cluster analysis and discriminant analysis to the example of the disease states diabetes type 1, diabetes type 2, and healthy. In a binary, supervised classification of any pair of these disease states we achieved specificities and sensitivities of approximately 80% within our data set.

Unilateral antegrade cerebral perfusion through the right axillary artery provides uniform flow distribution to both hemispheres of the brain: A magnetic resonance and histopathological study in pigs.

BACKGROUND: Bilateral antegrade cerebral perfusion (ACP) has decreased in popularity over the past decade because of its complexity and the risk of cerebral embolism. We used magnetic resonance (MR) perfusion imaging to assess flow distribution in both hemispheres of the brain during unilateral ACP through the right carotid artery via a cannula placed in the right axillary artery in conjunction with hypothermic circulatory arrest. METHODS AND RESULTS: Twelve pigs were randomly exposed to 120 minutes of either bilateral ACP through both carotid arteries (n=6) or unilateral ACP through the right axillary artery (n=6) at pressures of 60 to 65 mm Hg at 15 degrees C, followed by 60 minutes of cardiopulmonary bypass at 37 degrees C. MR perfusion images were acquired every 30 minutes before, during, and after ACP. The brain was perfusion fixed for histopathology. During initial normothermic cardiopulmonary bypass, MR perfusion imaging showed a uniform distribution of flow in the brain. In both the bilateral and unilateral ACP groups, the same pattern was maintained, with an increase in regional cerebral blood volume during ACP and reperfusion. The changes in regional cerebral blood volume and mean transit time were similar in both hemispheres during and after unilateral ACP. No difference was observed between the 2 groups. Histopathology showed normal morphology in all regions of the brain in both groups. CONCLUSIONS: Both bilateral ACP and unilateral ACP provide uniform blood distribution to both hemispheres of the brain and preserve normal morphology of the neurons after prolonged hypothermic circulatory arrest.

Classification of breast tumors by grade and steroid receptor status using pattern recognition analysis of infrared spectra.

Infrared (IR) spectroscopy applied to tissue sections yields complex spectra that provide a molecular fingerprint of the tissue. We have studied a cohort of 77 breast tumors by IR spectroscopy to develop an objective method for the assignment of grade of breast tumors. Although the major variations between spectra from different tumors were in absorptions arising from triglycerides (adipose tissue) and collagen, subtle changes in spectra could be detected that were independent of cellularity and tissue composition. Using a specific multivariate pattern recognition strategy to associate these changes in spectra with different tumor grades, we then were able to accurately reclassify tumors by grade (87% accuracy; kappa = 0.835). A similar approach allowed classification of steroid receptor status (93% accuracy; kappa = 0.852). We conclude that IR spectroscopy may have clinical utility in the objective assignment of breast tumor grade.

Retrograde cerebral perfusion results in flow distribution abnormalities and neuronal damage. A magnetic resonance imaging and histopathological study in pigs.

BACKGROUND: In the past few years, although significant efforts have been made to assess flow distribution during retrograde cerebral perfusion with microspheres, dye, or hydrogen clearance, flow distribution in real time is still undefined. We used MR perfusion imaging to monitor flow distribution in the brain during and after deep hypothermic circulatory arrest (DHCA) with antegrade or retrograde cerebral perfusion (ACP or RCP). METHODS AND RESULTS: Thirteen pigs were divided into 2 groups and exposed to 120 minutes of either RCP (n = 7) or ACP (n = 6) at 15 degrees C, followed by 60 minutes of cardiopulmonary bypass (CPB) at 37 degrees C. During DHCA, the brain was perfused antegradely through the common carotid artery or retrogradely through the superior vena cava at pressures of 60 to 70 mm Hg and 20 to 25 mm Hg in the ACP and RCP groups, respectively. Esophageal temperature was monitored continuously. MR perfusion images were acquired every 30 minutes before, during, and after DHCA. The brain was perfusion-fixed with formaldehyde solution for histopathology at the completion of each experiment. During initial normothermic CPB, MR perfusion imaging showed a nearly uniform distribution of flow in the brain. The same pattern was maintained with a significant increase in regional cerebral blood volume during ACP and reperfusion in the ACP group. RCP provided little or no detectable blood distribution to the brain, resulting in poor reperfusion of many areas of the brain on reflow with CPB at 37 degrees C. The total area suffering poor reperfusion was significantly higher in the RCP group than the ACP group. Histopathology showed no morphological changes in any area of the brain in the ACP group, whereas varying severity of neuronal damage was observed in different regions of the brain in the RCP group. CONCLUSIONS: ACP preserves uniform blood distribution and normal morphology of brain tissue after prolonged DHCA. RCP provides very little blood to the tissue of the brain. A 120-minute period of RCP results in abnormal flow distribution and neuronal damage during reperfusion. The damage resulting from shorter periods of RCP remains to be assessed.

Near-optimal region selection for feature space reduction: novel preprocessing methods for classifying MR spectra.

We introduce a global feature extraction method specifically designed to preprocess magnetic resonance spectra of biomedical origin. Such preprocessing is essential for the accurate and reliable classification of diseases or disease stages manifest in the spectra. The new method is genetic algorithm-guided. It is compared with our enhanced version of the standard forward selection algorithm. Both seek and select optimal spectral subregions. These subregions necessarily retain spectral information, thus aiding the eventual identification of the biochemistry of disease presence and progression. The power of the methods is demonstrated on two biomedical examples: the discrimination between meningioma and astrocytoma in brain tissue biopsies, and the classification of colorectal biopsies into normal and tumour classes. Both preprocessing methods lead to classification accuracies over 97% for the two examples.