Bartosz Swiderski

Texture characterization based on the Kolmogorov-Smirnov distance

We have developed new Kolmogorov-Smirnov method of description of the texture images.We have checked the performance of the proposed descriptor on the set of soil images.We have compared our solution to well known Haralick description of the texture. The paper proposes the new numerical descriptor of the texture based on the Kolmogorov-Smirnov (KS) statistical distance. In this approach to feature generation we consider the distribution of the pixel intensity placed in equal circular distances from the central point. In this statistical analysis each pixel of the image takes the role of the central point and KS statistics is estimated for the whole image. We determine the KS distance of pixel intensity corresponding to the coaxial rings of the increasing distance from the center. The slope of the linear regression function applied for approximating the characteristics presenting KS distance versus the geometrical distance of these rings, forms the proposed statistical descriptor of the image. We show the application of this numerical description for recognition of the set of images of soil of different type and show that it behaves very well as the diagnostic feature, better than texture Haralick features.

Multistage classification by using logistic regression and neural networks for assessment of financial condition of company

The paper presents the new approach to the automatic assessment of the financial condition of the company. We develop the computerized classification system applying WOE representation of data, logistic regression and Support Vector Machine (SVM) used as the final classifier. The applied method is a combination of a classical binary scoring approach and Support Vector Machine classification. The application of this method to the assessment of the financial condition of companies, classified into five classes, has shown its superiority with respect to classical approaches.

Aggregation of classifiers ensemble using local discriminatory power and quantiles

The paper presents a new approach to the dynamic classifier selection in an ensemble by applying the best suited classifier for the particular testing sample. It is based on the area under curve (AUC) of the receiver operating characteristic (ROC) of each classifier. To allow application of different types of classifiers in an ensemble and to reduce the influence of outliers, the quantile representation of the signals is used. The quantiles divide the ordered data into essentially equal-sized data subsets providing approximately uniform distribution of 0-1 support for each data point. In this way the recognition problem is less sensitive to the outliers, scales and noise contained in the input attributes. The numerical results presented for the chosen benchmark data-mining sets and for the data-set of images representing melanoma and non-melanoma skin lesions have shown high efficiency of the proposed approach and superiority to the existing methods. We developed new method of integrating classifiers in an ensemble based on quantiles.We have shown superiority of our solution on the benchmark problems.We have applied this solution to recognition of melanoma and proved its superiority.

Single-class SVM classifier for localization of epileptic focus on the basis of EEG

The paper presents the application of a single-class Support Vector Machine (SVM) for localization of the focus region at the epileptic seizure on the basis of EEG registration. The diagnostic features used in recognition are derived from the directed transfer function description, determined for different ranges of EEG signals. The results of the performed numerical experiments for the localization of the seizure focus in the brain have been confirmed by the real surgery of the brain for few patients.

False-positive reduction in computer-aided mass detection using mammographic texture analysis and classification

BACKGROUND AND OBJECTIVE The aim of computer-aided-detection (CAD) systems for mammograms is to assist radiologists by marking region of interest (ROIs) depicting abnormalities. However, the confusing appearance of some normal tissues that visually look like masses results in a large proportion of marked ROIs with normal tissues. This paper copes with this problem and proposes a framework to reduce false positive masses detected by CAD. METHODS To avoid the error induced by the segmentation step, we proposed a segmentation-free framework with particular attention to improve feature extraction and classification steps. We investigated for the first time in mammogram analysis, Hilberts image representation, Kolmogorov-Smirnov distance and maximum subregion descriptors. Then, a feature selection step is performed to select the most discriminative features. Moreover, we considered several classifiers such as Random Forest, Support Vector Machine and Decision Tree to distinguish between normal tissues and masses. Our experiments were carried out on a large dataset of 10168 ROIs (8254 normal tissues and 1914 masses) constructed from the Digital Database for Screening Mammography (DDSM). To simulate practical scenario, our normal regions are false positives asserted by a CAD system from healthy cases. RESULTS The combination of all the descriptors yields better results than each feature set used alone, and the difference is statistically significant. Besides, the feature selection steps yields a statistically significant increase in the accuracy values for the three classifiers. Finally, the random forest achieves the highest accuracy (81.09%), outperforming the SVM classifier (80.01%)) and decision tree (79.12%), but the difference is not statistically significant. CONCLUSIONS The accuracy of discrimination between normal and abnormal ROIs in mammograms obtained with the proposed gray level texture features sets are encouraging and comparable to these obtained with multiresolution features. Combination of several features as well as feature selection steps improve the results. To improve false positives reduction in CAD systems for breast cancer diagnosis, these features could be combined with multiresolution features.

Prediction of power consumption for small power region using indexing approach and neural network

The problem of prediction of 24-hour ahead power consumption in a small power region is a very important practical problem in power engineering. The most characteristic feature of the small region is large diversity of power consumption in the succeeding hours of the day making the prediction problem very hard. On the other side the accurate forecast of the power need for each of 24 hours of the next day enables to achieve significant saving on power delivery. The paper proposes the novel neural based method of forecasting the power consumption, taking into account the trend of its change associated with the particular hour of the day, type of the day as well as season of the year

Novel methods of image description and ensemble of classifiers in application to mammogram analysis

Abstract The paper proposes new advanced methods of image description and an ensemble of classifiers for recognition of mammograms in breast cancer. The non-negative matrix factorization and many other advanced methods of image representation, not exploited in the field of mammogram recognition, are developed and checked in the role of diagnostic features. Final image recognition is done by using an ensemble of classifiers. The new approach to the integration of an ensemble is proposed. It applies the weighted majority voting with the weights determined from the optimization task defined on the basis of the area under curve of ROC. The results of numerical experiments performed on large data base “Digital Database for Screening Mammography” containing more than 10,000 mammograms have confirmed superior accuracy in recognition of abnormal from the normal cases. The presented results of class recognition exceed the best achievements for this base reported in the actual publications.

Deep learning and non-negative matrix factorization in recognition of mammograms

This paper presents novel approach to the recognition of mammograms. The analyzed mammograms represent the normal and breast cancer (benign and malignant) cases. The solution applies the deep learning technique in image recognition. To obtain increased accuracy of classification the nonnegative matrix factorization and statistical self-similarity of images are applied. The images reconstructed by using these two approaches enrich the data base and thanks to this improve of quality measures of mammogram recognition (increase of accuracy, sensitivity and specificity). The results of numerical experiments performed on large DDSM data base containing more than 10000 mammograms have confirmed good accuracy of class recognition, exceeding the best results reported in the actual publications for this data base.

Deep learning in assessment of drill condition on the basis of images of drilled holes

This paper presents novel approach to drill condition assessment using deep learning. The assessment regarding level of the drill wear is done on the basis of the drilled hole images. Two states of the drill are taken into account: the sharp enough to continue production and worn out. The decision is taken on the basis of the shape of hole and also the level of hole shredding. In this way the drill condition is associated with the problem of image analysis and classification. Novel approach to this classification task in the form of deep learning has been applied in solving this problem. The important advantage of this method is great simplification of the recognition procedure, since any handy craft prepared features are not needed and the focus may be concentrated on the most interesting aspects of data mining and machine learning. The obtained results belong to the best in comparison to other approaches to the problem solution.

Nucleolus detection in the Fuhrman grading system for application in CCRC

Abstract The paper presents a method for nucleolus detection in images of nuclei in clear-cell renal carcinoma (CCRC). The method is based on the similarity of the nuclei image and the two-dimensional paraboloidal window function. The results of numerical experiments performed on almost 2600 images of CCRC nuclei have confirmed the good accuracy of the method. The developed algorithm will be used to accelerate further research in computer-assisted diagnosis of CCRC.