Publications

Early detection and diagnosis are critical factors to control the COVID-19 spreading. A number of deep learning-based methodologies have been recently proposed for COVID-19 screening in CT scans as a tool to automate and help with the diagnosis. These approaches, however, suffer from at least one of the following problems: (i) they treat each CT scan slice independently and (ii) the methods are trained and tested with sets of images from the same dataset. Treating the slices independently means that the same patient may appear in the training and test sets at the same time which may produce misleading results. It also raises the question of whether the scans from the same patient should be evaluated as a group or not. Moreover, using a single dataset raises concerns about the generalization of the methods. Different datasets tend to present images of varying quality which may come from different types of CT machines reflecting the conditions of the countries and cities from where they come from. In order to address these two problems, in this work, we propose an Efficient Deep Learning Technique for the screening of COVID-19 with a voting-based approach. In this approach, the images from a given patient are classified as group in a voting system. The approach is tested in the two biggest datasets of COVID-19 CT analysis with a patient-based split. A cross dataset study is also presented to assess the robustness of the models in a more realistic scenario in which data comes from different distributions. The cross-dataset analysis has shown that the generalization power of deep learning models is far from acceptable for the task since accuracy drops from 87.68% to 56.16% on the best evaluation scenario. These results highlighted that the methods that aim at COVID-19 detection in CT-images have to improve significantly to be considered as a clinical option and larger and more diverse datasets are needed to evaluate the methods in a realistic scenario.

A new mathematical programming formulation is proposed for an optimization problem in queueing networks. The sum of the blocking probabilities of a general service time, single server, finite, acyclic queueing network is minimized, as are the total butter sizes and the overall service rates. A multi-objective genetic algorithm (MOGA) and a particle swarm optimization (MOPSO) algorithm are combined to solve this difficult stochastic problem. The derived algorithm produces a set of efficient solutions for multiple objectives in the objective function. The implementation of the optimization algorithms is dependent on the generalized expansion method (GEM), a classical tool used to evaluate the performance of finite queueing networks. A set of computational experiments is presented to attest to the efficacy and efficiency of the proposed approach. Insights obtained from the analysis of a complex network may assist in the planning of these types of queueing networks.

We present a robustness analysis of an inter-cities mobility complex network, motivated by the challenge of the COVID-19 pandemic and the seek for proper containment strategies. Brazilian data from 2016 are used to build a network with more than five thousand cities (nodes) and twenty-seven states with the edges representing the weekly flow of people between cities via terrestrial transports. Nodes are systematically isolated (removed from the network) either at random (failures) or guided by specific strategies (targeted attacks), and the impacts are assessed with three metrics: the number of components, the size of the giant component, and the total remaining flow of people. We propose strategies to identify which regions should be isolated first and their impact on people mobility. The results are compared with the so-called reactive strategy, which consists of isolating regions ordered by the date the first case of COVID-19 appeared. We assume that the nodes' failures abstract individual municipal and state initiatives that are independent and possess a certain level of unpredictability. Differently, the targeted attacks are related to centralized strategies led by the federal government in agreement with municipalities and states. Removing a node means completely restricting the mobility of people between the referred city/state and the rest of the network. Results reveal that random failures do not cause a high impact on mobility restraint, but the coordinated isolation of specific cities with targeted attacks is crucial to detach entire network areas and thus prevent spreading. Moreover, the targeted attacks perform better than the reactive strategy for the three analyzed robustness metrics.

Multimodal systems are a workaround to enhance the robustness and effectiveness of biometric systems. A proper multimodal dataset is of the utmost importance to build such systems. The literature presents some multimodal datasets, although, to the best of our knowledge, there are no previous studies combining face, iris/eye, and vital signals such as the Electrocardiogram (ECG). Moreover, there is no methodology to guide the construction and evaluation of a chimeric dataset. Taking that fact into account, we propose to create a chimeric dataset from three modalities in this work: ECG, eye, and face. Based on the Doddington Zoo criteria, we also propose a generic and systematic protocol imposing constraints for the creation of homogeneous chimeric individuals, which allow us to perform a fair and reproducible benchmark. Moreover, we have proposed a multimodal approach for these modalities based on state-of-the-art deep representations built by convolutional neural networks. We conduct the experiments in the open-world verification mode and on two different scenarios (intra-session and inter-session), using three modalities from two datasets: CYBHi (ECG) and FRGC (eye and face). Our multimodal approach achieves impressive decidability of 7.20 ± 0.18, yielding an almost perfect verification system (i.e., Equal Error Rate (EER) of 0.20% ± 0.06) on the intra-session scenario with unknown data. On the inter-session scenario, we achieve a decidability of 7.78 ± 0.78 and an EER of 0.06% ± 0.06. In summary, these figures represent a gain of over 28% in decidability and a reduction over 11% of the EER on the intra-session scenario for unknown data compared to the best-known unimodal approach. Besides, we achieve an improvement greater than 22% in decidability and an EER reduction over 6% in the inter-session scenario.

For multi-objective optimization algorithms, one of the key features is the ability to find a good approximation to the optimal trade-off among the multiple objectives. The most traditional algorithms focus only on the distribution of solutions in the objective space. Nevertheless, a good representation of solutions in the decision space is also important from the point of view of the decision-making process. This work presents a dominance-weighted uniformity-based multi-objective algorithm. The selection phase of the algorithm performs a greedy search of the uniformity measure weighted by a certain measure of dominance. Preliminary experiments suggest that this method promotes the uniformity of the population in the decision variable space while keeping the convergence performance in the objective space.

AbstractThe spatial scan statistic is a widely used technique for detecting spatial clusters. Several extensions of this technique have been developed over the years. The objectives of these techniques are the detection accuracy improvement and a flexibilization on the search clusters space. Based on Voronoi-Based Scan (VBScan), we propose a biobjective approach using a recursively VBScan method called multi-objective multiple clusters VBScan (MOMC-VBScan), alongside a new measure called matching. This approach aims to identify and delineate all multiple significant anomalies in a search space. We conduct several experiments on different simulated maps and two real datasets, showing promising results. The proposed approach proved to be fast and with good precision in determining the partitions.

The QRS complex is a very important part of a heartbeat in the electrocardiogram signal, and it provides useful information for physicians to diagnose heart diseases. Accurately detecting the fiducial points that compose the QRS complex is a challenging task. Another issue concerning the QRS detection is its computational costs since the algorithm should have a fast and real-time response. In this context, there is a trade-off between computational cost and precision. Convolutional networks are a deep learning approach, and it has achieved impressive results in several computer vision and pattern recognition problems. Nowadays there is hardware that fully embeds convolutional network models, significantly reducing computational cost for real-world and real-time applications. In this direction, this work proposes a deep learning approach, based on convolutional network, aiming to detect heartbeat pattern. We tested two different architectures with two different proposes, one very deep and that has small receptive fields, and the other that has larger receptive fields. Preliminary experiments on the MIT-BIH arrhythmia database showed that the studied convolutional network presents promising results for QRS detection which are comparable with state-of-the-art methods.

In this work, we propose to detect the iris and periocular regions simultaneously using coarse annotations and two well-known object detectors: YOLOv2 and Faster R-CNN. We believe coarse annotations can be used in recognition systems based on the iris and periocular regions, given the much smaller engineering effort required to manually annotate the training images. We manually made coarse annotations of the iris and periocular regions (≈122K images from the visible (VIS) spectrum and ≈38K images from the near-infrared (NIR) spectrum). The iris annotations in the NIR databases were generated semi-automatically by first applying an iris segmentation CNN and then performing a manual inspection. These annotations were made for 11 well-known public databases (3 NIR and 8 VIS) designed for the iris-based recognition problem, and are publicly available to the research community1. Experimenting our proposal on these databases, we highlight two results. First, the Faster R-CNN + Feature Pyramid Network (FPN) model reported an Intersection over Union (IoU) higher than YOLOv2 (91.86% vs 85.30%). Second, the detection of the iris and periocular regions being performed simultaneously is as accurate as performed separately, but with a lower computational cost, i.e. two tasks were carried out at the cost of one.

The iris is considered as the biometric trait with the highest unique probability. The iris location is an important task for biometrics systems, affecting directly the results obtained in specific applications such as iris recognition, spoofing and contact lenses detection, among others. This work defines the iris location problem as the delimitation of the smallest squared window that encompasses the iris region. In order to build a benchmark for iris location we annotate (iris squared bounding boxes) four databases from different biometric applications and make them publicly available to the community. Besides these 4 annotated databases, we include 2 others from the literature. We perform experiments on these six databases, five obtained with near infra-red sensors and one with visible light sensor. We compare the classical and outstanding Daugman iris location approach with two window based detectors: 1) a sliding window detector based on features from Histogram of Oriented Gradients (HOG) and a linear Support Vector Machines (SVM) classifier; 2) a deep learning based detector fine-tuned from YOLO object detector. Experimental results showed that the deep learning based detector outperforms the other ones in terms of accuracy and runtime (GPUs version) and should be chosen whenever possible.

The spatial scan statistic is widely used by public health professionals in the detection of spatial clusters in inhomogeneous point process. The most popular version of the spatial scan statistic uses a circular-shaped scanning window. Several other variants, using other parametric or non-parametric shapes, are also available. However, none of them offer information about the uncertainty on the borders of the detected clusters.

The global expansion of biometric systems promotes the emergence of new and more robust biometric modalities. In that context, electroencephalogram (EEG) based biometric interest has been growing in recent years. In this study, a novel approach for EEG representation, based on deep learning, is proposed. The method was evaluated on a database containing 109 subjects, and all 64 EEG channels were used as input to a Deep Convolution Neural Network. Data augmentation techniques are explored to train the deep network and results showed that the method is a promising path to represent brain signals, overcoming baseline methods published in the literature.

Usually, in the deep learning community, it is claimed that generalized representations that yielding outstanding performance / effectiveness require a huge amount of data for learning, which directly affect biometric applications. However, recent works combining transfer learning from other domains have surmounted such data application constraints designing interesting and promising deep learning approaches in diverse scenarios where data is not so abundant. In this direction, a biometric system for the periocular region based on deep learning approach is designed and applied on two non-cooperative ocular databases. Impressive representation discrimination is achieved with transfer learning from the facial domain (a deep convolutional network, called VGG) and fine tuning in the specific periocular region domain. With this design, our proposal surmounts previous state-of-the-art results on NICE (mean decidability of 3.47 against 2.57) and MobBio (equal error rate of 5.42% against 8.73%) competition databases.

This paper presents a hybrid PSO algorithm (Particle Swarm Optimization) with an ILS (Iterated Local Search) operator for handling equality constraints problems in mono-objective optimization problems. The ILS can be used to locally search around the best solutions in some generations, exploring the attraction basins in small portions of the feasible set. This process can compensate the difficulty of the evolutionary algorithm to generate good solutions in zero-volume regions. The greatest advantage of the operator is the simple implementation. Experiments performed on benchmark problems shows improvement in accuracy, reducing the gap for the tested problems.

There are several biometric-based systems which rely on a single biometric modality, most of them focus on face, iris or fingerprint. Despite the good accuracies obtained with single modalities, these systems are more susceptible to attacks, i.e, spoofing attacks, and noises of all kinds, especially in non-cooperative (in-the-wild) environments. Since non-cooperative environments are becoming more and more common, new approaches involving multi-modal biometrics have received more attention. One challenge in multimodal biometric systems is how to integrate the data from different modalities. Initially, we propose a deep transfer learning optimized from a model trained for face recognition achieving outstanding representation for only iris modality. Our feature level fusion by means of features selection targets the use of the Particle Swarm Optimization (PSO) for such aims. In our pool, we have the proposed iris fine-tuned representation and a periocular one from previous work of us. We compare this approach for fusion in feature level against three basic function rules for matching at score level: sum, multi, and min. Results are reported for iris and periocular region (NICE.II competition database) and also in an open-world scenario. The experiments in the NICE.II competition databases showed that our transfer learning representation for iris modality achieved a new state-of-the-art, i.e., decidability of 2.22 and 14.56% of EER. We also yielded a new state-of-the-art result when the fusion at feature level by PSO is done on periocular and iris modalities, i.e., decidability of 3.45 and 5.55% of EER.

This paper presents a hybrid PSO algorithm with a VND-based operator for handling equality constraint problems in continuous optimization. The VND operator can be defined both as a local search and a kind of elitism operator for equality constraint problems playing the role of ``fixing'' the best estimates of the feasible set. Experiments performed on benchmark problems suggest that the VND operator can enhance both the convergence speed and the accuracy of the final result.

Classifying arrhythmias can be a tough task for a human being and automating this task is highly desirable. Nevertheless fully automatic arrhythmia classification through Electrocardiogram (ECG) signals is a challenging task when the inter-patient paradigm is considered. For the inter-patient paradigm, classifiers are evaluated on signals of unknown subjects, resembling the real world scenario. In this work, we explore a novel ECG representation based on vectorcardiogram (VCG), called temporal vectorcardiogram (TVCG), along with a complex network for feature extraction. We also fine-tune the SVM classifier and perform feature selection with a particle swarm optimization (PSO) algorithm. Results for the inter-patient paradigm show that the proposed method achieves the results comparable to state-of-the-art in MIT-BIH database (53% of Positive predictive (+P) for the Supraventricular ectopic beat (S) class and 87.3% of Sensitivity (Se) for the Ventricular ectopic beat (V) class) that TVCG is a richer representation of the heartbeat and that it could be useful for problems involving the cardiac signal and pattern recognition.

The automation process of Pap smear analysis holds the potential to address women's health care in the face of an increasing population and respective collected data. A fundamental step for automating analysis is cell detection from light microscopy images. Such information serves as input to cell classification algorithms and diagnostic recommendation tools. This paper describes an approach to nuclei cell segmentation, which critically impacts the following steps for cell analyses. We developed an algorithm combining clustering and genetic algorithms to detect image regions with high diagnostic value. A major problem when performing the segmentation of images is the cellular overlay. We introduce a new nuclear targeting approach using heuristics associated with a multi-objective genetic algorithm. Our experiments show results using a public 45-image dataset, including comparison to other cell detection approaches. The findings suggest an improvement in the nuclei segmentation and promise to support more sophisticated schemes for data quality control.

For decades iris recognition has been widely studied by the scientific community due to its almost unique and stable patterns. Iris recognition biometric systems apply mathematical pattern-recognition techniques to an iris' image of an individual's eye to extract its feature vector. Comparing the dissimilarities from two feature vectors with an acceptance threshold, the system decides if the two vectors are from the same individual's eye. If applied in a well-controlled environment, iris recognition can achieve outstanding accuracies, however, to accomplish that in non-controlled environments is still a challenge researchers are constantly trying to compensate open issues in this context. In order to better explore the patterns found in the iris, researchers have recently begun using a classification approach using multiple signatures, hoping to improve the algorithm robustness. This work aims to explore the effectiveness and scalability of using multiple signatures with a 2-D Gabor filter in a biometric verification system through iris recognition. This is done using two independent Genetic Algorithms to search for the best parameters to the feature extraction technique and on the acceptance frontier search. The method was evaluated by analyzing the behavior of the Half Total Error Rate (HTER) when the number of partitions varies. The experiments showed that the best result was found with 12 partitions on the iris, reaching 0.21% of HTER.