Horacio M. González-Velasco

Neural-networks-based edges selector for boundary extraction problems

In the present work, a neural-networks-based system is presented that makes it possible to reduce, when generating edge maps to be used in an object boundary detection problem, the number of edges that are not due to the object itself, but to the background. Starting from a conventional edge detection, the selection is carried out by a neuralnetworks-based classifier, which is trained using examples. As a test for the system, the application to bovine livestock images is presented, from which we want to extract the boundary of the animal.

Consistent performance measurement of a system to detect masses in mammograms based on blind feature extraction

BackgroundBreast cancer continues to be a leading cause of cancer deaths among women, especially in Western countries. In the last two decades, many methods have been proposed to achieve a robust mammography‐based computer aided detection (CAD) system. A CAD system should provide high performance over time and in different clinical situations. I.e., the system should be adaptable to different clinical situations and should provide consistent performance.MethodsWe tested our system seeking a measure of the guarantee of its consistent performance. The method is based on blind feature extraction by independent component analysis (ICA) and classification by neural networks (NN) or SVM classifiers. The test mammograms were from the Digital Database for Screening Mammography (DDSM). This database was constructed collaboratively by four institutions over more than 10 years. We took advantage of this to train our system using the mammograms from each institution separately, and then testing it on the remaining mammograms. We performed another experiment to compare the results and thus obtain the measure sought. This experiment consists in to form the learning sets with all available prototypes regardless of the institution in which them were generated, obtaining in that way the overall results.ResultsThe smallest variation from comparing the results of the testing set in each experiment (performed by training the system using the mammograms from one institution and testing with the remaining) with those of the overall result, considering the success rate for an intermediate decision maker threshold, was roughly 5%, and the largest variation was roughly 17%. But, if we considere the area under ROC curve, the smallest variation was close to 4%, and the largest variation was about a 6%.ConclusionsConsidering the heterogeneity in the datasets used to train and test our system in each case, we think that the variation of performance obtained when the results are compared with the overall results is acceptable in both cases, for NN and SVM classifiers. The present method is therefore very general in that it is able to adapt to different clinical situations and provide consistent performance.

Study of a mammographic CAD performance dependence on the considered mammogram set

This work analyzes the influence of the set of mammograms used in the training processes of a computer aided diagnosis system on the overall performance. We used the mammograms provided by the Digital Database for Screening Mammography, one of the most extended research database. The obtained results seem to suggest an effect on the performance values obtained in a CAD system with different database subsets. Therefore, in order to make valid comparisons between CAD systems, the specification of the mammogram set used to test the system is of the utmost importance.

Independent Component Analysis for Cloud Screening of Meteosat Images

In this work we use Independent Component Analysis (ICA) as feature extraction stage for cloud screening of Meteosat images covering the Iberian Peninsula. The images are segmented in the classes land (L), sea (S), fog (F), low clouds (CL), middle clouds (CM), high clouds (CH) and clouds with vertical growth (CV). The classification of the pixels of the images is performed with a back propagation neural network (BPNN) from the features extracted by applying the FastICA algorithm over 3x3, 5x5 and 7x7 pixel windows of the images.

Independent component analysis applied to detection of early breast cancer signs

This work evaluates the efficiency of Independent Component Analysis in conjunction with neural network classifiers to detect microcalcification clusters in digitized mammograms, the most important non invasive sign of breast cancer. The widespread Digital Database for Screening Mammography was used as the source for digitized mammograms. The results seem to suggest that this technique is suitable to deal with the noisy mammogram environment.A method for providing real-time weather and other information about an airport to an aircraft. The method comprises the steps of generating a signal representing instantaneous weather information, determining a real-time weather information value from the instantaneous weather information signal over a predetermined time interval, generating an address signal representing the real-time weather information value, storing a plurality of signals representing real-time weather information messages, generating one of the message signals corresponding to the real-time weather information value in response to the address signal, and transmitting the message signal to a receiver on the aircraft.

ICA and GA feature extraction and selection for cloud classification

In this work we tackle a particular case of image segmentation, the automatic detection of the amount and type of clouds over the Iberian Peninsula using satellite images. To segment the images we classify each pixel of the image into one of the classes defined using a neural network and a set of features representative of the pixel. We emphasized in the preprocessing stage, extracting and selecting a suitable set of features from the images to carry out an optimal classification. To carry out the feature extraction we use the independent component analysis (ICA) algorithm. The features extracted with this algorithm are very dependent on the dimension of the patches, so we extract several sets of features, one for each value of the dimension of the patch. All of these sets of features are joined together to form an initial characteristic vector of the pixels of the images. Finally, we reduce the dimensionality of this initial characteristic vector by means of Genetic Algorithms (GA), choosing the best subset of features that offer the best classification results.In this work we tackle a particular case of image segmentation, the automatic detection of the amount and type of clouds over the Iberian Peninsula using satellite images. To segment the images we classify each pixel of the image into one of the classes defined using a neural network and a set of features representative of the pixel. We emphasized in the preprocessing stage, extracting and selecting a suitable set of features from the images to carry out an optimal classification. To carry out the feature extraction we use the independent component analysis (ICA) algorithm. The features extracted with this algorithm are very dependent on the dimension of the patches, so we extract several sets of features, one for each value of the dimension of the patch. All of these sets of features are joined together to form an initial characteristic vector of the pixels of the images. Finally, we reduce the dimensionality of this initial characteristic vector by means of Genetic Algorithms (GA), choosing the best subset of features that offer the best classification results.

Using GAs to obtain an optimal set of codes for an ultrasonic local positioning system

Signal coding and pulse compression techniques have been recently introduced in Local Positioning Systems as a means to enhance the measurement precision of these systems and to increase their operation frequency. This work presents a Genetic Algorithm that performs the search for an optimal family of binary codes, to be used in a system based on ultrasonic technology. The developed algorithm takes into account the transduction effect of the emitters on the correlation properties of the modulated family to obtain a set of codes that exhibits a superior performance than other families previously used.

A method for interactive shape detection in cattle images using genetic algorithms

Segmentation methods based on deformable models have proved to be successful with difficult images, particularly those using genetic algorithms to minimize the energy function. Nevertheless, they are normally conceived as fully automatic, and not always generate satisfactory results. In this work, a method to include the information of fixed points whithin a contour detection system using point distribution models and genetic algorithms is presented. Also, an interactive scheme is proposed to take advantage of this technique. The method has been tested against a database of 93 cattle images, with a significant improvement in the success rate of the detections, from 61% up to 95%.

Semi-automatic Measure and Identification of Allergenic Airborne Pollen

Current lifestyle in developed countries makes the practice of outdoor activities to be almost mandatory. But, since these practices such as trekking, biking, horseback, or simply running or walking in urban parks, are made in nature (at least outdoors) not everyone can practice them in optimal physical conditions at any time of the year. We are referring to those who suffer from pollinosis or “hay fever”.

Robustness of a CAD system on digitized mammograms

In this paper we study the robustness of our CAD system, since this is one of the main factors that determine its quality. A CAD system must guarantee consistent performance over time and in various clinical situations. Our CAD system is based on the extraction of features from the mammographic image by means of Independent Component Analysis, and machine learning classifiers, such as Neural Networks and Support Vector Machine. To measure the robustness of our CAD system we have used the digitized mammograms of the USFs DDSM database, because this database was built by digitizing mammograms from four different institutions (four different scanner) during more than 10 years. Thus, we can use the mammograms digitized with one scanner to train the system and the remaining to evaluate the performance, what gives us a measure of the robustness of our CAD system.