Nicandro Cruz-Ramírez

Cervical cancer detection using colposcopic images: a temporal approach

In the present work, we propose a methodology analysis of the colposcopic images to help the expert to make a more robust diagnosis of precursor lesions of cervical cancer. Although some others approaches have been used to assess cervical lesion, a complete methodology to evaluate temporal changes of tissue color is still missing. The different processes involved in the analysis are described. The image registration was implemented using the phase correlation method followed by a locally applied algorithm based on the normalized cross-correlation. During the parameterization process, each time series obtained from the image sequences was represented as a parabola in a parameter space. A supervised Bayesian learning approach is proposed to classify the features in the parameter space according to the classification made by the colposcopist. Then those labels are used as a criterion to categorize the tissue and perform the image segmentation. Some preliminary results are shown using unsupervised learning with real data.

Discovering interobserver variability in the cytodiagnosis of breast cancer using decision trees and Bayesian networks

We evaluate the performance of two decision tree procedures and four Bayesian network classifiers as potential decision support systems in the cytodiagnosis of breast cancer. In order to test their performance thoroughly, we use two real-world databases containing 692 cases and 322 cases collected by a single observer and 19 observers, respectively. The results show that, in general, there are considerable differences in all tests (accuracy, sensitivity, specificity, PV+, PV- and ROC) when a specific classifier uses the single-observer dataset compared to those when this same classifier uses the multiple-observer dataset. These results suggest that different observers see different things: a problem known as interobserver variability. We graphically unveil such a problem by presenting the structures of the decision trees and Bayesian networks resultant from running both databases.

Aceto-white temporal pattern classification using k-NN to identify precancerous cervical lesion in colposcopic images

After Pap smear test, colposcopy is the most used technique to diagnose cervical cancer due to its higher sensitivity and specificity. One of the most promising approaches to improve the colposcopic test is the use of the aceto-white temporal patterns intrinsic to the color changes in digital images. However, there is not a complete understanding of how to use them to segment colposcopic images. In this work, we used the classification algorithm k-NN over the entire length of the aceto-white temporal pattern to automatically discriminate between normal and abnormal cervical tissue, reaching a sensitivity of 71% and specificity of 59%.

Application of time series discretization using evolutionary programming for classification of precancerous cervical lesions

In this work, we present a novel application of time series discretization using evolutionary programming for the classification of precancerous cervical lesions. The approach optimizes the number of intervals in which the length and amplitude of the time series should be compressed, preserving the important information for classification purposes. Using evolutionary programming, the search for a good discretization scheme is guided by a cost function which considers three criteria: the entropy regarding the classification, the complexity measured as the number of different strings needed to represent the complete data set, and the compression rate assessed as the length of the discrete representation. This discretization approach is evaluated using a time series data based on temporal patterns observed during a classical test used in cervical cancer detection; the classification accuracy reached by our method is compared with the well-known times series discretization algorithm SAX and the dimensionality reduction method PCA. Statistical analysis of the classification accuracy shows that the discrete representation is as efficient as the complete raw representation for the present application, reducing the dimensionality of the time series length by 97%. This representation is also very competitive in terms of classification accuracy when compared with similar approaches.

Differential evolution with combined variants for dynamic constrained optimization

In this work a differential evolution algorithm is adapted to solve dynamic constrained optimization problems. The approach is based on a mechanism to detect changes in the objective function and/or the constraints of the problem so as to let the algorithm to promote the diversity in the population while pursuing the new feasible optimum. This is made by combining two popular differential evolution variants and using a memory of best solutions found during the search. Moreover, random-immigrants are added to the population at each generation and a simple hill-climber-based local search operator is applied to promote a faster convergence to the new feasible global optimum. The approach is compared against other recently proposed algorithms in an also recently proposed benchmark. The results show that the proposed algorithm provides a very competitive performance when solving different types of dynamic constrained optimization problems.

Bayes-N: An Algorithm for Learning Bayesian Networks from Data Using Local Measures of Information Gain Applied to Classification Problems

Bayes-N is an algorithm for Bayesian network learning from data based on local measures of information gain, applied to problems in which there is a given dependent or class variable and a set of independent or explanatory variables from which we want to predict the class variable on new cases. Given this setting, Bayes-N induces an ancestral ordering of all the variables generating a directed acyclic graph in which the class variable is a sink variable, with a subset of the explanatory variables as its parents. It is shown that classification using this variables as predictors performs better than the naive bayes classifier, and at least as good as other algorithms that learn Bayesian networks such as K2, PC and Bayes-9. It is also shown that the MDL measure of the networks generated by Bayes-N is comparable to those obtained by these other algorithms.

How good are the bayesian information criterion and the minimum description length principle for model selection? a bayesian network analysis

The Bayesian Information Criterion (BIC) and the Minimum Description Length Principle (MDL) have been widely proposed as good metrics for model selection. Such scores basically include two terms: one for accuracy and the other for complexity. Their philosophy is to find a model that rightly balances these terms. However, it is surprising that both metrics do often not work very well in practice for they overfit the data. In this paper, we present an analysis of the BIC and MDL scores using the framework of Bayesian networks that supports such a claim. To this end, we carry out different tests that include the recovery of gold-standard network structures as well as the construction and evaluation of Bayesian network classifiers. Finally, based on these results, we discuss the disadvantages of both metrics and propose some future work to examine these limitations more deeply.

An Agents and Artifacts Approach to Distributed Data Mining

This paper proposes a novel Distributed Data Mining (DDM) approach based on the Agents and Artifacts paradigm, as implemented in CArtAgO [9], where artifacts encapsulate data mining tools, inherited from Weka, that agents can use while engaged in collaborative, distributed learning processes. Target hypothesis are currently constrained to decision trees built with J48, but the approach is flexible enough to allow different kinds of learning models. The twofold contribution of this work includes: i) JaCA-DDM: an extensible tool implemented in the agent oriented programming language Jason [2] and CArtAgO [10,9] to experiment DDM agent-based approaches on different, well known training sets. And ii) A collaborative protocol where an agent builds an initial decision tree, and then enhances this initial hypothesis using instances from other agents that are not covered yet (counter examples); reducing in this way the number of instances communicated, while preserving accuracy when compared to full centralized approaches.

A parsimonious constraint-based algorithm to induce Bayesian network structures from data

In this paper, we present a novel algorithm, called MP-Bayes, which induces Bayesian network structures from data based on entropy measures. One of the main features of this method is its parsimonious nature: it tends to represent the joint probability distribution underlying the data with the least number of arcs. While other methods that build Bayesian networks tend to overfit the data, MP-Bayes creates models that seem to have an adequate trade-off between accuracy and complexity. To support such a claim, we compare the performance of MP-Bayes, in terms of classification, against those of four different Bayesian network classifiers. The results show that our procedure generalizes well in a wide range of situations.

A Repair Method for Differential Evolution with Combined Variants to Solve Dynamic Constrained Optimization Problems

Repair methods, which usually require feasible solutions as reference, have been employed by Evolutionary Algorithms to solve constrained optimization problems. In this work, a novel repair method, which does not require feasible solutions as reference and inspired by the differential mutation, is added to an algorithm which uses two variants of differential evolution to solve dynamic constrained optimization problems. The proposed repair method replaces a local search operator with the aim to improve the overall performance of the algorithm in different frequencies of change in the constrained space. The proposed approach is compared against other recently proposed algorithms in an also recently proposed benchmark. The results show that the proposed improved algorithm outperforms its original version and provides a very competitive overall performance with different change frequencies.