• Integrating fNIRS and machine learning: Shedding a light on Parkinson's disease detection

The purpose of this research is to introduce an approach to assist the diagnosis of Parkinson’s disease (PD) by classifying functional near infra-red spectroscopy (fNIRS) studies as PD positive or negative. fNIRS is a non-invasive optical signal modality that conveys the brain’s hemodynamic response, specifically changes in blood oxygenation in the cerebral cortex; and its potential as a tool to assist PD detection deserves to be explored since it is non-invasive and cost-effective as apposed to other neuroimaging modalities. Besides the integration of fNIRS and machine learning, a contribution of this work is that various approaches were implemented and tested to find the implementation that achieves the highest performance. All the implementations used a logistic regression model for classification. A set of 792 temporal and spectral features were extracted from each participant’s fNIRS study. In the two best performing implementations, an ensemble of feature-ranking techniques was used to select a reduced feature subset, which was subsequently reduced with a genetic algorithm. The best detection performance achieved was accuracy of 100%, precision of 100 %, recall of 100 %, F1 score of 1, area under the curve of 1 by using 14 features.

• Detection and classification of skin burns on color images using multi-resolution clustering and the classification of reduced feature subsets

The detection of skin burns on color images and their corresponding classification to identify their intensity degree are useful tools for automatic assistance in emergencies in which clinical experience is not available. In this work, the analysis of images with skin burns starts with multiresolution processing to identify non-homogeneous regions that contain these lesions. Subsequently, segmentation within the non-homogeneous region is performed with k-means so that these regions are divided into unlabeled clusters. Finally, each unclassified cluster is assigned one of six classes; healthy skin, first-degree burn, second-degree burn, third-degree burn, shadowed skin, background. Four classification models are used; k-nearest neighbors, multi-layer perceptron, linear discriminant analysis, and support vector machine. There is a very large amount of available features that can be associated with skin analysis, but only a few of them are substantially associated with the classification of skin as being burnt or not, and with the burn degree. A contribution of this work is the selection of relevant features for each of two tasks: (1) segmentation of the regions of interest into clusters and (2) classification of each cluster to detect and classify burns. According to the selected features, the best detection and classification performance metrics are 95.65% sensitivity and 94.83% precision.

• Prediction of epileptic seizures based on convolutional neural networks (CNN) and functional Near-Infra Red Spectroscopy (fNIRS) signals

This work presents the implementation of the learning algorithm for a Deep Learning Machine (DLM), specifically, a Convolutional Neural Network (CNN), which is effectively applied to the prediction of epileptic seizures from functional Near-Infrared Spectroscopy (fNIRS) recordings, an optical modality for recording of brain waves. A CNN is suitable for this application, instead of other learning machines, since fNIRS recordings are characterized by very high dimensionality, given that there are hundreds of fNIRS channels and that at any time position these signals are represented as two-dimensional tensors. The training of the CNN is based on the back propagation algorithm to update kernel weights in convolutional layers and synaptic weights in classification layers. This work presents the required formulations to update weights over all the layers of the network. Application of CNN to fNIRS recordings showed an accuracy ranging between 96.9% and 100%, depending on the subject. The most important aspect of obtaining these results is the combination of fNIRS signals with the particular CNN algorithm. This signal modality has not been used in epileptic seizure prediction and this work pretends to be one of the first to use this modality and deep learning to address the problem of seizure prediction.

Collaboration with Dr. Edgar Guevara (CONACYT - UASLP), Dr. Frédéric Lesage (Ecole Polytechnique de Montréal), Dr. Philippe Pouliot (Ecole Polytechnique de Montréal), Dr. Dang Khoa Nguyen (Hôpital Notre-Dame du CHUM), Dr. Ke Peng (École Polytechnique Montréal de Montréal).

• Transiting exo-planet identification using machine learning techniques

Spatial missions such as the Kepler mission, and the Transiting Exoplanet Survey Satellite (TESS) mission, have encouraged data scientists to analyze light curve datasets. The purpose of analyzing these data is to look for planet transits, with the aim of discovering and validating exoplanets, which are planets found outside our Solar System. Furthermore, transiting exoplanets can be better characterized when light curves and radial velocity curves are available. The manual examination of these datasets is a task that requires big quantities of time and effort, and therefore is prone to errors. As a result, the application of machine learning methods has become more common on exoplanet discovery and categorization research. This research focuses on the analysis on different exoplanet transit discovery algorithms based on machine learning, some of which even found new exoplanets. The analysis of these algorithms is organized in four steps, namely light curve preprocessing, possible exoplanet signal detection, and identification of the detected signal to decide whether it belongs to an exoplanet or not. This work concludes that multiresolution analysis in the time-frequency domain can improve exoplanet signal identification, because of the characteristics of light curves and transiting exo-planet signals.

• Fully automatic alpha matte extraction in green screen images using artificial neural networks

The alpha matte is a two-dimensional map that is used to combine two images, one containing a foreground and the other containing a background, a process known as digital compositing. Alpha matte extraction is performed on green-screen images and requires user interaction to tune parameters in different pre-processing and post-processing stages to refine an alpha matte. This work tackles the problem of fully automatic extraction of the foreground on a green screen image along with the extraction of the corresponding alpha matte, a process also known as pulling a matte. The method is based on learning machine that assigns an alpha value, from a discrete set of ten alpha values, to each patch on a green-screen image. The approach for assigning an alpha value to a patch is based on a set of features that enhance discrimination between foreground and background. Prior to alpha matte extraction, the classifier is trained to learn to separate foreground objects from green screen backgrounds as well as to generate the corresponding alpha matte map required for subsequent digital compositing. We tested the proposed method on high-definition (HD) green screen images, corresponding to two different sequence cases (TOY CAR ROTO & GODIVA MEDIUM), both cases from a public data set provided by Hollywood Camera Work LLC. HD green screen images, from each particular case, have limitations that are not good enough for film production. In addition, a data set with 64 images was generated to test how the proposed approach handles alpha matte extraction under unsuitable conditions such as short separation distance between the subject and the green screen. The project also provides, for the case of the public data set, a quantitative comparison between the alpha matte, extracted by means of the proposed approach, and that generated by the application Adobe After Effects CC, which has the disadvantage of demanding a large amount of user interaction. 

Collaboration with Omar López-Rincón & Dr. Oleg Starostenko-Basarab (UDLAP).

• Forecasting of stock return prices with sparse representation of financial time series over redundant dictionaries 

This project introduces the theory, methodology and application of a new predictive model for time series within the financial sector, specifically data from 20 companies listed on the U. S. stock exchange market. The proposed method is based on learned redundant dictionaries for sparse representation of financial time series. The methodology is conducted by finding the optimal set of atoms for the predicting model following two directions for the generation of dictionaries; by extraction of atoms from past daily return price values to build untrained dictionaries and by atom extraction followed by training of dictionaries through K-SVD (K - Singular Value Decomposition). Prediction of financial time series is a periodic process where each cycle consists of two stages: (1) training of the model to learn the dictionary that maximizes the probability of occurrence of an observation sequence of return values, (2) prediction of the return value for the next coming trading day. After prediction, the two stages are alternatively repeated by using an adjusted sequence of observations that adds the newest observed return value and drops the oldest observed return value from the sequence (window shift). The motivation for such research is the fact that a tool, which might generate confidence of the potential benefits obtained from using formal financial services, would encourage more participation in a formal system such as the stock market. Theory, issues, challenges and results related to the application of sparse representation to the prediction of financial time series, as well as the performance of the method, are presented.

• A method to assist in the diagnosis of early diabetic retinopathy: Image processing applied to detection of microaneurysms in fundus images

Diabetes increases the risk of developing any deterioration in the blood vessels that supply the retina, an ailment known as Diabetic Retinopathy (DR). Since this disease is asymptomatic, it can only be diagnosed by an ophthalmologist. However, the growth of the number of ophthalmologists is lower than the growth of the population with diabetes so that preventive and early diagnosis is difficult due to the lack of opportunity in terms of time and cost. Preliminary, affordable and accessible ophthalmological diagnosis will give the opportunity to perform routine preventive examinations, indicating the need to consult an ophthalmologist during a stage of non-proliferation. During this stage, there is a lesion on the retina known as microaneurysm (MA), which is one of the first clinically observable lesions that indicate the disease. In recent years, different image processing algorithms, which allow the detection of the DR, have been developed; however, the issue is still open since acceptable levels of sensitivity and specificity have not yet been reached, preventing its use as a pre-diagnostic tool. Consequently, this work introduces a new approach for MA detection based on (1) reduction of non-uniform illumination; (2) normalization of image grayscale content to improve dependence of images from different contexts; (3) application of the bottom-hat transform to leave reddish regions intact while suppressing bright objects; (4) binarization of the image of interest with the result that objects corresponding to MAs, blood vessels, and other reddish objects (Regions of Interest - ROIs) are completely separated from the background; (5) application of the hit-or-miss Transformation on the binary image to remove blood vessels from the ROIs; (6) two features are extracted from a candidate to distinguish real MAs from FPs, where one feature discriminates round shaped candidates (MAs) from elongated shaped ones (vessels) through application of Principal Component Analysis (PCA); (7) the second feature is a count of the number of times that the radon transform of the candidate ROI, evaluated at the set of discrete angle values {0°, 1°, 2°, …, 180°}, is characterized by a valley between two peaks. The proposed approach is tested on the public databases DiaretDB1 and Retinopathy Online Challenge (ROC) competition. The proposed MA detection method achieves sensitivity, specificity and precision of 92.32%, 93.87% and 95.93% for the diaretDB1 database and 88.06%, 97.47% and 92.19% for the ROC database. Theory, results, challenges and performance related to the proposed MA detecting method are presented.

Collaboration with Dr. Jorge Martínez-Carballido (INAOE).

• Remote detection of forest fires from video signals with classifiers based on K-Singular Value Decomposition (K-SVD) learned dictionaries

In this project a method for remote detection of forest fires in video signals from surveillance cameras is introduced. The idea is based on learned redundant dictionaries for sparse representation of feature vectors extracted from image patches on three different regions; smoke, sky and ground. A testing image patch is assigned to the region for which the corresponding dictionary gives the best sparse representation during segmentation. To further reduce the presence of misclassified patches, a spatio-temporal cuboid of patches is built around a classified patch to take a majority vote in the set of classes inside the cuboid. To reduce the number of false positives there is a verification process to determine if a region of interest is growing. Theory, results, issues and challenges related to the implementation of the forest fire monitoring system, and performance of the method are presented.

• Segmentation of endocardium in ultrasound images based on sparse representation over learned redundant dictionaries 

This project tackles the problem of segmenting the endocardium in 2-D short-axis echocardiographic images from rats by using the sparse representation of feature vectors over learned dictionaries during classification. We highlight important aspects of the application of the theory of sparse representation and dictionary learning to the problem of ultrasound image segmentation. Experiments were conducted following two directions for the generation of dictionaries for myocardium and blood pool regions; by manual extraction of image patches to build untrained dictionaries and by patch extraction followed by training of dictionaries. The results obtained from different learned dictionaries are compared. During classification of an image patch, instead of using features of the patch alone, features of neighboring patches are combined.

Collaboration with Dr. Hemant D. Tagare (Yale University).

• A human action recognition approach with a reduced feature set based on the natural domain knowledge of the human figure 

Current video surveillance systems are not designed to raise an automatic alert in case of situations that put people lives at risk such as accidents, assaults and terrorism among others. This is due to the fact that these systems are not able to analyze huge amounts of video signals at higher processing speed where these signals come from cameras installed in the worldwide network. Faced with this situation, scientific communities are combining efforts to design algorithms and hardware to accelerate the processing of video signals. However, most of the methods proposed to date are too complex to be implemented in hardware at the place where the video camera is installed. In this paper, we report a significantly reduced feature set to design an analysis algorithm of significant less complexity which recognizes human actions from video sequences. The proposed method is based on the natural domain knowledge of the human figure such as proportions of the human body and foot positions. The analysis is characterized by working on sub-sequences of the entire video signals, processing a small fragment of the whole image, estimating the location of the region of interest, using simple operations (sum, subtraction, multiplications, divisions), extracting a reduced number of features per frame (6 features), and using a combination of four linear classifiers (one perceptron and three support vector machines) with a hierarchical structure. The method is evaluated on two of the datasets cited in the human action recognition literature, the Weizmann and the UIUC datasets. Results show that for the case of the Weizmann dataset, the correct classification rate (CCR) is 99.95% when the LOOCV Protocol is used and 98.38% for the case of Protocol 60-40, which is comparable or even higher than that of current state-of-the-art methods. Confusion matrices were also obtained for the UIUC dataset, where the obtained CCR is 100% for the case of the LOOCV Protocol and 99.35% when Protocol 60-40 is used. The experimental results are promising with much fewer features (between 85 and 113 times less features), compared with other methods, and the possibility of processing more than 200 fps.

Collaboration with Dr. Jorge Martínez-Carballido (INAOE).

• False positive reduction by an annular model as a set of few features for microcalcification detection to assist early diagnosis of breast cancer

Early automatic breast cancer detection from mammograms is based on the extraction of lesions, known as microcalcifications (MCs). This research proposes a new and simple system for microcalcification detection to assist in early breast cancer detection. This work uses the two most recognized public mammogram databases, MIAS and DDSM. The MC detection method is based on (1) Beucher gradient for detection of regions of interest (ROIs), (2) an annulus model for extraction of few and effective features from candidates to MCs, and (3) one classification stage with two different classifiers, k Nearest Neighbor (KNN) and Support Vector Machine (SVM). For dense mammograms in the MIAS database, the performance metrics achieved are sensitivity of 0.9835, false alarm rate of 0.0083, accuracy of 0.9835, and area under the ROC curve of 0.9980 with a KNN classifier. The proposed MC detection method, based on a KNN classifier, achieves, a sensitivity, false positive rate, accuracy and area under the ROC curve of 0.9813, 0.0224, 0.9795 and 0.9974 for the MIAS database; and 0.9035, 0.0439, 0.9298 and 0.9759 for the DDSM database. By slightly reducing the true positive rate the method achieves three instances with false positive rate of 0:2 on fatty mammograms with KNN and SVM, and one on dense with SVM. The proposed method gives better results than those from state of the art literature, when the mammograms are classified in fatty, fatty-glandular, and dense.

Collaboration with Dr. Jorge Martínez-Carballido (INAOE).