Abstract
In this paper we derive an upper bound for the average-case generalization error of the mixture of experts modular neural network, based on an average-case generalization error bound for an isolated neural network. By doing this we also generalize a previous bound for this architecture that was restricted to special problems. We also present a correction factor for the original average generalization error, that was empirically obtained, that yields more accurate error bounds for the 6 data sets used in the experiments. These experiments illustrate the validity of the derived error bound for the mixture of experts modular neural network and show how it can be used in practice.

Abstract
We present a fully automated method for extracting the lung region from volumetric X-ray CT images based on material decomposition. By modeling the human thorax as a composition of different materials, the proposed method follows a threshold-based, hierarchical voxel classification strategy. The segmentation procedure involves the automatic computation of threshold values and consists on three main steps: patient segmentation and decomposition, large airways extraction and lung parenchyma decomposition, and lung region of interest segmentation. Experimental results were performed on thoracic. CT images acquired from 30 patients. The method provides a reproducible set of thresholds for accurate extraction of the lung parenchyma, needed for computer aided diagnosis systems.

Abstract
The aim of this paper is to examine a multiclassifier approach to the classification of the lung nodules in X-ray chest radiographs. The approach investigated here is based on an image region-based classification whose output is the information of the presence or absence of a nodule in an image region. The classification was made, essentially, in two steps: firstly, a set of rotation invariant features was extracted from the responses of a multi-scale and multi-orientation filter bank; secondly, different classifiers (multi-layer perceptrons) are designed using different features sets and trained in different data. These classifiers are further combined in order to improve the classification performance. The obtained results are promising and can be used for reducing the false-positives nodules detected in a computer-aided diagnosis system.

Abstract
The paper presents the methodology developed to solve the class imbalanced problem that occurs in the classification of Thin-Layer Chromatography (TLC) images. The proposed methodology is based on resampling, and consists in the undersampling of the majority class (normal class), while the minority classes, which contain Lysosomal Storage Disorders (LSD) samples, are oversampled with the generation of synthetic samples. For image classification two approaches are presented, one based on a hierarchical classifier and another uses a multiclassifier system, where both classifiers are trained and tested using balanced data sets. The results demonstrate a better performance of the multiclassifier system using the balanced sets.

Abstract
In spite of significant advances in image segmentation techniques, evaluation of these methods thus far has been largely subjective. Typically, the effectiveness of a new algorithm is demonstrated only by the presentation of a few segmented images that axe evaluated by some method, or it is otherwise left to subjective evaluation by the reader. We propose a new approach for evaluation of segmentation that takes into account not only the accuracy of the boundary localization of the created segments but also the under-segmentation and over-segmentation effects, regardless to the number of regions in each partition. In addition, it takes into account the way humans perceive visual information. This new metric can be applied both to automatically provide a ranking among different segmentation algorithms and to find an optimal set of input parameters of a given algorithm.

Abstract
Grouping and segmentation of images remains a challenging problem in computer vision. Recently, a number of authors have demonstrated a good performance on this task using spectral methods that are based on the eigensolution of a similarity matrix. In this paper, we implement a variation of the existing methods that combines aspects from several of the best-known eigenvector segmentation algorithms to produce a discrete optimal solution of the relaxed continuous eigensolution.