Perfusion Quantification using Gaussian Processes

I.K. Andersen, A. Szymkowiak, C. Rasmussen,
L.K. Hansen, J.R. Marstrand, H.B.W. Larsson

Proceedings of the joint meeting of the International Society of Magnetic Resonance in Medicine and the European Society of Magnetic Resonance in Medicine and Biology, 2001

Abstract: The Gaussian Process has not previously been used for deconvolution with application to quantifying perfusion form the dynamic R2* measurements. The Gaussian Process is compared to the Singular Value Decomposition methods currently used, and is found to be superior.

Automated malignant melanoma diagnosis: Neural network analysis of chemical alterations presented in Raman spectra of the cancer

Abstract: Malignant melanoma (MM) is the most aggressive skin cancer. Early detection and removal of MM is curative, however the specificity and sensitivity of clinical diagnosis is low and reach only about 80% even in specialised centres. Here, we investigated whether the chemical changes in the melanoma tissue can be detected by Raman spectroscopy and used for tumor diagnosis. Near-infrared Fourier transform (NIR-FT) Raman spectra were obtained of samples of malignant melanoma (MM), the lesions from other skin disorders which can clinically be confused with MM: pigmented nevi (NV), basal cell carcinoma (BCC), seborrhoeic keratosis (SK) and normal skin (NOR). The spectra were analysed visually and by neural network software. Interestingly, the sensitivity analysis of the spectral frequencies used by neural network for MM classification picked up independently frequencies previously used on visual inspection. We propose that neural network analysis of NIR-FT Raman spectra could constitute a novel method for rapid, automated skin cancer diagnosis.

On supervised learning with labeled and unlabeled data

L.K. Hansen

Submitted April 2001

Abstract: Based on an analysis of the average generalization error for
conditional predictive distributions I conclude that transductive Bayesian learning is generalization optimal for realizable learning problems with a well-specified prior. In particular, all available data including unlabeled data should be used in forming the likelihood. However, for the unrealizable case the situation is non-trivial, asymptotic results indicate that unlabeled data should be used with care, corroborating empirical
results obtained recently by Nigam et al.

Webmining: Learning from the World Wide Web

J. Larsen, L.K. Hansen, A. Szymkowiak, T. Christiansen and T. Kolenda

To appear in special issue of Computational Statistics and Data Analysis, 2001 and Proceedings of Nonlinear Methods and Data Mining 2000, Rome, Italy, Sept. 25-26, 2000, pp. 106-125.

Abstract: These papers study the use of advanced datamining algorithms for mining webpages. These algorithms are widely applicable to medical processing systems and telemedicine and adding flexibilty, effectiveness and user firendliness. Automated analysis of the world wide web is a new challenging area relevant in many applications, e.g., retrieval, navigation and organization of information, automated information assistants, and e-commerce. This paper discusses the use of unsupervised and supervised learning methods for user behavior modeling and content-based segmentation and classification of web pages. The modeling is based on independent component analysis and hierarchical probabilistic clustering techniques.

[paper (pdf)]

A non-parametric model for separation of histograms

G. Maletti

Danish Conference on Pattern Recognition (DANCOMB'00). Aalborg, Denmark, 31st August - 1st September 2000.

Abstract: The objective of the work is to separate histograms of classes present in a given image. It is done without setting any parameters. The hypothesis is that information about each pixel of the image can be obtained from its neighbors. The idea is to use a variable resolution
sensory system as an image data reduction scheme. In the present work we associate to each pixel the highest resolution (an optimum window) by detecting among its neighbors the emergence of a state of greater order with an increasing function of the fraction of examples and by maximizing the redundancy in the data model. Each pixel is associated to a statistic associated to its optimum window. The output of this model is a new image with well-separated histograms of the classes.

Active Learning from the Context Applied to Diffusion Reduction

G. Maletti, B. Ersbøll and K. Conradsen

Submitted for the 3rd International Conference on Modeling and Using Context (CONTEXT'01), 2001.

Contextual Supervised Classifier Based on Normalized Histograms

G. Maletti, B. Ersbøll and K. Conradsen

Submitted for the 11th International Conference on Image Analysis an Processing (ICIAP), 2001.

An Initial Training Set Generation Scheme

Abstract:

Supervised classifiers employ a-priori information of each determined class. This is usually obtained by means of training sets interactively generated or with semi-automatic schemes that still depend on some user input. The purpose of this work is to generate an initial statistically valid sample of a class starting from only one single prototype pixel. This is done by means of a new mapping model that uses redundant contextual information. The initial sample generated is the maximum size
neighborhood centered on the prototype pixel, which contains pixels that belong to the same class as the center one. This is defined by a window of optimum size obtained by minimizing the variance of a learning variable over the line of maximal redundancy of the mapping model. The minimum of the function studied corresponds to the estimation of the point in which a new class emerges. Additionally, a supervised classifier was designed and implemented as an application of the model proposed.

A supervised contextual classification scheme based on an
automated region growing algorithm.

Abstract: A supervised classification scheme to segment optical multi-band images has been developed. In this classifier, an automated region-growing algorithm delineates the training sets. This algorithm handles three parameters: an initial pixel seed, a window and a threshold for each class. A suitable pixel seed is manually implanted through visual inspection of the image classes. The optimum value for the window and the threshold are obtained from spectral distances. These distances are calculated from mathematical models of spectral separabilities. A pixel is incorporated into a region if a spectral homogeneity criterion is satisfied in the pixel-centered window
for a given threshold. The homogeneity criterion is obtained from the models of spectral distances. The set of pixels forming a region represents a statistically valid sample of a defined class signaled by the initial pixel seed. The grown regions constitute therefore optimum training sets for each class. Comparing the statistical behavior of a sliding window with that of each class performs the classification. For classification, a set of windows is used that provides the best separability among the classes. The centered pixel of the sliding window is labeled as belonging to a class if its spectral distance is a minimum to the class. A series of examples, employing synthetic images are presented to show the value of this classifier. The goodness of the segmentation is evaluated by means of the Kappa coefficient, and a matrix of distances derived from the model of spectral separabilities.

On Comparison of Adaptive Regularization Methods

S. Sigurdsson, J. Larsen and L.K. Hansen

Proceedings of the IEEE Workshop on Neural Networks for Signal Processing X, B. Widrow, L. Guan, K. Paliwa, T. Adali, J. Larsen, E. Wilson, S. Douglas (eds.), Piscataway, New Jersey: IEEE. Sydney, Australia, Dec. 11-13, 2000, pp. 221-230.

Abstract: Neural networks are important models for diagnosis support and has been used for classification of skin lesions. In this study we address the control of model complexity and generalization ability which finds expression in the ubiquitous bias-variance dilemma. Regularization is a tool for optimizing the model structure reducing variance at the expense of introducing extra bias. The overall objective of adaptive regularization is to tune the amount of regularization ensuring minimal generalization error. This paper investigates recently suggested adaptive regularization schemes. Some methods focus directly on minimizing an estimate of the generalization error (either algebraic or empirical), whereas others starts from different criteria, e.g., the Bayesian evidence. We suggest various algorithm extensions and performed numerical experiments with linear models.

[paper (pdf)]

Detection of Skin Cancer by Classification of Raman Spectra

S. Sigurdsson, P.A. Philipsen, L.K. Hansen, J. Larsen, M. Gniadecka, H.C. Wulf.

In preparation April 2001

Abstract: Skin lesion classification based on Raman spectroscopy is approached using linear and neural network non-linear classifiers. Special attention is devoted to visualization of the learned models. Resampling experiments allow us to estimate the so-called learning curve, i.e., the generalization performance as function of the sample size. We find the typical generalization cross-over, namely that for small sample size, a linear model perform better than a more complex neural net model. As sample size increases eventually the non-linear model prevails.
The best performance for the present dataset involving 150 cases and 5 classes, is 80% correct, comparable to expert classifications based on visual inspection.

[paper (pdf)]

Hierachical Clustering for Datamining

A. Szymkoviak, J. Larsen, L.K. Hansen

In preparation, May 2001

Abstract: This work present a hierarchical clustering techniques based on Gaussian mixture models for datamining relevant for medical diagnosis support and telemedicine. The method features optimal selection of subspace dimension, automatic determination of number of cluster, techniqes for interpretation of clusters, detection of outliers and the use of labeled and unlabeled examples. The methods is applied to email segmentation.

Imputating Missing Values in Diary Records of Sun-Exposure Study

A. Szymkowiak, P.A. Philipsen, J. Larsen, L.K. Hansen, E. Thieden, H.C. Wulf

In preparation April 2001

Abstract: In a sun-exposure study, questionnaires concerning sun-habits were collected from 195 subjects (the group of people involved in the experiment). Before processing this data, the general problem
with missing values has to be faced. They occur when some, or even all, the questions have not been answered in a questionnaire. In this paper, the missing values of low concentration are investigated. We suggest two different models for interpolating, namely the statistical Gaussian model and the non-parametric K-nearest Neighbor model.

[paper (pdf)]