An evolutionary Michigan recurrent fuzzy system for nuclei classification in cytological images using nuclear chromatin distribution.

OBJECTIVE: The objective of this research is to carry out the classification of cellular nuclei in cytological pleural fluid images. The article focuses on the feature extraction and classification processes. The extracted feature is a spatial measurement of the chromatin distribution in cellular nuclei. The designed classifiers are fuzzy classifiers that carry out supervised classification. The classifier system's inputs are data series that represent these texture measurements. METHODS AND MATERIAL: The classifier is built on a Recurrent Fuzzy System (RFS). An evolutionary algorithm inspired by the Michigan approach is used to find an optimal RFS to classify different patterns expressed as data series. RESULTS: The effectiveness of the proposed classifier system is compared with other existing classification methods and evaluated via Receiver Operating Characteristic (ROC) analysis. We have obtained RFS based classifiers that perform with sensitivity values between 82.26 and 93.55% and with specificity values between 80.65 and 90.32%. The behavior of the proposed chromatin measurement is also compared with other texture measurements. CONCLUSION: The RFS based classifiers were successfully applied to the proposed data series that represent the chromatin distribution in cellular nuclei. These fuzzy classifiers present the highest classification efficiency and the ROC analysis confirms their suitable behavior.

Exploiting the advantages of symbolically interpretable continuous biomedical parameters with a fuzzyfied symbolic model.

Continuous biomedical parameters are normally hybrid signals, because they contain both sub-symbolic and symbolic information. This paper describes a methodology to design adequate processing systems for the automatic analysis of this kind of signals. The methodology is supported by the concept of fuzzy system.

Automatic analysis of signals with symbolic content.

This paper presents a set of methods for helping in the analysis of signals with particular features that admit a symbolic description. The methodology is based on a general discrete model for a symbolic processing subsystem, which is fuzzyfied by means of a fuzzy inference system. In this framework a number of design problems have been approached. The curse of dimensionality problem and the specification of adequate membership functions are the main ones. In addition, other strategies, which make the design process simpler and more robust, are introduced. Their goals are automating the production of the rule base of the fuzzy system and composing complex systems from simpler subsystems under symbolic constrains. These techniques are applied to the analysis of wakefulness episodes in the sleep EEG. In order to solve the practical difficulty of finding remarkable situations from the outputs of the symbolic subsystems an unsupervised adaptive learning technique (FART network) has been applied.

Tools for acquisition, processing and knowledge-based diagnostic of the electroencephalogram and visual evoked potentials.

The objective of our research is to develop computer-based tools to automate the clinical evaluation of the electroencephalogram (EEG) and visual evoked potentials (VEP). This paper describes a set of solutions to support all the aspects regarding the standard procedures of the diagnosis in neurophysiology, including: (1) acquisition and real-time processing and compression of EEG and VEP signals, (2) real-time brain mapping of spectral powers, (3) classifier design, (4) automatic detection of morphologies through supervised neural networks. (5) signal analysis through fuzzy modelling, and (6) a knowledge based approach to classifier design.