Catálogo de publicaciones - libros

The presented method of extraction of the shape of the microcalcifications (MCs) is based on the watershed segmentation (WS) and region merging. Assuming that the locations of the MCs are known, we use two kinds of markers for the WS: the internal marker indicating the interior of the MC, and the external marker separating the MCs. Carrying out the WS in the area between these two markers we obtain a limited number of regions. These regions are merged into a mask of an MC by maximization of the average contrast between the mask and its surroundings. The obtained shapes of the masks agree with human intuition and can be used for the classification of MCs as to their malignancy.

In this paper, the possibilities of using ultrasonic projection for computer recognition of biological structures were analyzed. The computer simulation of the distribution of ultrasonic wave propagation velocity local values inside two three-dimensional objects dipped in the water was done, and next for a few projection planes the Radon transforms were calculated in the form of the mean value distribution in parallel geometry. The results of simulation were visualized in the form of the grey scale images and in the form of the pseudo-3D charts with lighting. The real measurements of three-dimensional biological objects were carried out on the elaborated research setup. The obtained projection images of the distribution of three different acoustic parameters in the object structure: propagation velocity, amplitude and mid frequency shift of ultrasonic wave were compared with the optical scan in the same plane of object cross-section. The conclusion is that computer-assisted ultrasonic projection enables the correct recognition of biological structures. On the basis of a few projection images of the examined structure obtained from many directions it is possible to attempt to computer-reconstruct in 3D heterogeneity boundaries inside the structure, which can be used e.g. in diagnosing the early stages of women’s breast cancer.

The main goal of our research was attempt to answer the question, if machine learning algorithms could be applied in computer aided medical diagnostics. Executed tests proven, that data mining methods based on machine learning can be used in medical diagnostics, but it can not substitute an expert, especially in case of rare diseases.

Classifiers induced from revised datasets have better classification accuracy. It is indicating that quality of training data has significant influence for induced classifier accuracy.

In expert opinion of expert, the most of obtained decision rules are consistent with medical knowledge. All cases of incorrect classification were caused by insufficient mathematical model. The evaluation of exact mathematical model in medicine is very difficult.

In this paper results of experiments on two popular machine learning algorithms were presented. The appliance of other classification methods, like Bayesian classifiers, neural networks and fuzzy sets, is subject of future research.

Most of reports on computer supported cytological investigations focus on searching for objective, quantitative descriptors enabling an automated system to distinguish between “normal” and “pathological” objects, usually cells or their organelles. A great number of sophisticated tools have been developed and reported. However, few reports may be found concerning the problem of detecting artefacts in cytological smears and reducing their influence on the overall system performance. On the other hand, the problem is crucial for the whole system setup and if not properly solved may spoil any attempts to implement the system in practice. The paper addresses this neglected problem trying to point out some general rules and procedures that should be followed to reject artefacts from automatic cytological analysis.

Quantitative perfusion imaging using Dynamic Susceptibility Contrast (DSC) MRI method requires to measure the Arterial Input Function (AIF) and deconvolve it from the measured tissue signal. We present a method for automatic recognition of the global AIF based on multistage algorithm. The method is validated using real world (clinically measured) DSC-MRI image series. Only 5% of all automatically generated AIFs (one series) were rejected by the expert. The method can be easily extended to produce a set of local AIFs and can be used as fully automatic or as an intelligent assistant tool for a neuroradiologist.

In this paper the new method for the automatic segmentation and classification of fundus eye images taken from classical fundus camera into normal and glaucomatous ones is proposed. The presented method consists of the following three stages: segmentation, feature selection, and classification. The mean sensitivity of the proposed method is 93%, while the mean specificity is 97%.

In the future, voice messages in natural language will undoubtedly be the most important way of communication between humans and machines. Great progress is made in many fields of science, where communication between the technological devices and the operator is an important task, e.g. motorization, road traffic, etc. The condition of the effectiveness of the presented intelligent two-way voice communication system between the technological device and the operator is to equip it with mechanisms of command verification and correctness. In the automated processes of production, the condition for safe communication between the operator and the technological device is analyzing the state of the technological device and the process before the command is given and using artificial intelligence for assessment of the technological effects and safety of the command. In operations of the automated technological processes, many process states and various commands from the operator to the technological device can be distinguished. A large number of combined technological systems characterize the realization of that process. In complex technological processes, if many parameters are controlled, the operator is not able to analyze a sufficient number of signals and react by manual operations on control buttons. The research aiming at developing an intelligent layer of two-way voice communication is very difficult, but the prognosis of the technology development and its first use shows a great significance in efficiency of supervision and production humanization.

The paper considers recognition of isolated words of the Polish Sign Language using a canonical stereo system that observes the signer from a frontal view. Recognition is based on human skin detection and Hidden Markov Models. Several feature vectors taking into account information about the hand shape and 3D position of the hand with respect to the face are examined. To improve the recognition rate the classifiers are combined by voting or by fuzzy integral. We focus on 101 words that can be used at the doctor’s and at the post office.

The paper presents author’s method of shape evaluation adapted to quality estimation of characters printed by typewriter. The shown method is based on statistical analysis of the Maximal Square Map (MSM) described in detail in the paper. This method has been elaborated for quality evaluation of computer aided information retrieval from archival machine typed paper documents.

An approach to solving the problem of the high-quality system Text-to-Speech (TTS) for Polish language synthesis is considered in this paper. Synthesis of phonemic speech characteristics is based on Polish language linguistic resources analysis and Allophones Natural Waves (ANW) method of speech signal concatenation.