Christian Pellegrini

Colour displays and look-up tables: Real time modification of digital images

Image processing in biomedical research has become customary, along with use of colour displays to run image processing packages. The performance of softwares is highly dependent on the device they run on: architecture of colour display, depth of frame buffer, existence of look-up table, etc. Knowledge of such basic features is therefore becoming very important, especially because results can differ from device to device. This introductory paper discusses hardware features and software applications. A general architecture of colour displays is exposed, comparing the features of the most commonly used devices. Basic organisation of memory, electron gun and screen are analysed for each type of display, concluding with a more detailed study of raster scan devices. Frame buffer and look-up table organisation are then analysed in relation with overhead expenses such as time and memory. Relation between image data and displayed images is discussed. By means of examples, the manipulation of colour tables is examined in detail, showing how to improve display of images without altering image data. Finally, the basic operations performed by the look-up table editor developed at University of Geneva are presented.

One-Class Support Vector Machines with a Conformal Kernel. A Case Study in Handling Class Imbalance

Class imbalance is a widespread problem in many classification tasks such as medical diagnosis and text categorization. To overcome this problem, we investigate one-class SVMs which can be trained to differentiate two classes on the basis of examples from a single class. We propose an improvement of one-class SVMs via a conformal kernel transformation as described in the context of binary SVM classifiers by [2,3]. We tested this improved one-class SVM on a health care problem that involves discriminating 11% nosocomially infected patients from 89% non infected patients. The results obtained are encouraging: compared with three other SVM-based approaches to coping with class imbalance, one-class SVMs achieved the highest sensitivity recorded so far on the nosocomial infection dataset. However, the price to pay is a concomitant decrease specificity, and it is for domain experts to decide the proportion of false positive cases they are willing to accept in order to ensure treatment of all infected patients.

Constraining the MLP power of expression to facilitate symbolic rule extraction

Extracting symbolic rules from multilayer perceptrons is an important open question, especially when input neurons are continuous. To solve this problem we constrain the power of expression of a standard MLP with threshold functions in the hidden layer. In this case, hyper-plane equations are precisely determined and translated into symbolic rules. We illustrate our interpretable MLP (IMLP) in two applications; one from iris classification, and one from coronary heart disease diagnosis. In spite of the reduced power of expression, IMLP is able to give close mean predictive accuracy with respect to a standard MLP.

An architecture for musical score recognition using high-level domain knowledge

Proposes an original approach to musical score recognition, a particular case of high-level document analysis. In order to overcome the limitations of existing systems, we propose an architecture which allows for a continuous and bidirectional interaction between high-level knowledge and low-level data, and which is able to improve itself over time by learning. This architecture is made of three cooperating layers, one made of parameterized feature detectors, another working as an object-oriented knowledge repository and the other as a supervising Bayesian metaprocessor. Although the implementation is still in progress, we show how this architecture is adequate for modeling and processing knowledge.

Intelligent Assistance for Coronary Heart Disease Diagnosis: A Comparison Study

Using only non invasive medical information, we propose inductive decision trees exploiting C4.5 algorithm, artificial neural networks with three MLP models, and linear discriminant analysis to diagnose coronary heart disease. The first neural network model is a constructive MLP called OIL (Orthogonal Incrementing Learning) that builds its hidden neurons during the training phase. The second one is a fixed MLP architecture with the same number of hidden neurons obtained from the first network building methodology. The last one is a special ”interpretable” MLP model with a fixed architecture (IMLP), which is interpretable through symbolic rule extraction. In general, explanation of connectionist model responses are difficult to obtain, especially when input examples have continuous variables. This is not acceptable for real world diagnosis applications. The novelty in our study consists in the interpretability of the IMLP model we have developed. For this diagnosis application, all neural networks globally obtain better predictive accuracies than C4.5 and the linear discriminant analysis. Results obtained with the OIL method are slightly better than those obtained by IMLP, but they lack interpretability.

A metalevel architecture for knowledge-based neural network design

In this paper, the term knowledge-based neural network (NN) design is used to refer to all efforts at exploiting prior knowledge in neural network configuration and training. A variety of techniques have been proposed for this purpose; SCANDAL provides a workbench for evaluating and integrating these techniques. After a quick overview of three main approaches to NN design, we describe SCANDALS multi-agent, metalevel architecture as well as its strategies for maximizing the use of domain knowledge. To assess the impact of prior knowledge on NN performance, experiments were conducted comparing knowledge-based techniques with a search-based configuration algorithm. Results show that the use of prior knowledge in neural network design leads to both faster learning and improved generalization. More interestingly, this appears to hold even when domain knowledge and data are deficient; in such cases, knowledge is extracted from the available data and is used both to configure the network and to generate artificial training instances. This leads us to hope that time-consuming iterative search can be avoided even in knowledge-lean domains.

HIGH ORDER STATISTICS FOR IMAGE CLASSIFICATION

This paper addresses the problem of image classification using local information which is aggregated to provide global representation of different image classes. Local information is adaptively extracted from an image database using Independent Component Analysis (ICA) which provides a set of localized, oriented, and band-pass filters selective to independent features of the images. Local representation using ICA techniques has been previously investigated by several researchers. However, very little work has been done on further use of these representations to provide more complex and global description of images. In this paper, we present an algorithm which uses the energy of a minimal set of ICA filters to provide class-specific signatures which are shown to be strongly discriminant. Computer simulations are carried on two image databases, one consisting of five classes--referred to as categories--(buildings, rooms, mountains, forests and beaches) and one consisting of a set of 30 objects from multiple views for viewpoint invariant object recognition. The classification performance of the algorithm using both Independent and Principal Component Analyses are reported and discussed.

Accurate Decomposition of Standard MLP Classification Responses into Symbolic Rules

In this work we determine hyper-plane equations from three MLP models. The first one is the standard MLP model, the second one is called OMLP (oblique MLP) and the last one is called IMLP (Interpretable MLP). From OMLP and IMLP, hyper-plane equations are determined easily, whereas for MLP we just give a sufficient condition for the detection of potential hyper-plane discriminators.

Automatic' learning strategies and their application to electrophoresis analysis

Automatic learning plays an important role in image analysis and pattern recognition. A taxonomy of automatic learning strategies is presented; this categorization is based on the amount of inferences the learning element must perform to bridge the gap between environmental and system knowledge representation level. Four main categories are identified and described: rote learning, learning by deduction, learning by induction, and learning by analogy. An application of learning by induction to medical image analysis is then exposed. It consists in the classification of two-dimensional gel electrophoretograms into meaningful distinct classes, as well in their conceptual description.

Recognizing Images from ICA Filters and Neural Network Ensembles with Rule Extraction

This work presents ensembles of neural network models that learn to discriminate images from different categorical scenes. The basic idea was to use ICA filter energies and to train neural network ensem- bles. The presented results improved the predictive accuracy of previ- ously published work on the second classification problem. Finally, rules generated from ensembles in the less complex classification task showed that a few filters are sufficient to reach a good recognition rate, whereas many more filters are represented in the rule antecedents of the most difficult classification problem.