Felipe da Rocha Henriques

A interpretação de ultrassonografias renais depende de inspeção visual subjetiva e da delimitação anatômica do rim. Este artigo propõe um pipeline de visão computacional que primeiro segmenta o rim e depois analisa regiões hiperecogênicas suspeitas próximas às pirâmides renais. Os modelos foram retreinados com as imagens originais da base kidneyUS e aplicação de CLAHE no pré-processamento. Em um teste deduplicado de 70 imagens, a U-Net atingiu Dice de 0,9290, IoU de 0,8722 e precisão de 0,9236, obtendo o maior Dice entre as arquiteturas avaliadas. Aplicado a 4.479 imagens externas, esse modelo gerou 3.534 pseudomáscaras candidatas, ainda pendentes de revisão humana e ausentes das métricas apresentadas.

The design of Wireless Sensor Networks (WSNs) involves balancing conflicting objectives related to sensing quality, connectivity, and resource consumption. This becomes more challenging in heterogeneous networks, where sensor nodes differ in capability, coverage, and cost. This paper addresses the Sensor Allocation Problem through a multi-objective integer linear programming formulation. The proposed model considers a set of predefined candidate locations for node deployment and multiple transceiver types with different communication ranges and energy consumption characteristics, using data from commercial devices. An e-constraint approach is used to minimize total average energy consumption subject to minimum coverage requirements, thereby generating Pareto-efficient network topologies. The model is evaluated through 225 computational experiments with different monitoring-region sizes and candidate-location densities. The results highlight that the proposed formulation can generate feasible and connected heterogeneous WSN topologies and capture the trade-off between coverage requirements and energy consumption.

Este trabalho propõe um modelo de rede neural convolucional 3D derivado da arquitetura R3D_18 para classificar imagens de ressonância magnética (MRI) do joelho, diferenciando imagens normais e anormais associadas à osteoartrite. Utilizando o dataset OAI-MRI-3DDESS, o modelo foi treinado com validação cruzada e oversampling para compensar o desbalanceamento de classes. A abordagem visa aprimorar a precisão diagnóstica e reduzir a sobrecarga dos profissionais, explorando as informações volumétricas das imagens. Os experimentos indicaram desempenho robusto, com acurácia superior a 86% e AUC próximo a 0,92, mostrando o potencial do método para aplicações clínicas.

Fraudes em transações com cartões de crédito são um desafio global, resultando em grandes prejuízos financeiros. Este trabalho propõe um simulador de dados sintéticos de transações para replicar a dinâmica de dados reais. Esses dados foram usados para criar modelos baseados em algoritmos de classificação e detecção de anomalias, capazes de identificar fraudes. Desafios como modelagem sequencial, mudança de contexto, feedback atrasado e peculiaridades dos dados foram abordados. O algoritmo Random Forest destacou-se, detectando 76,7% das fraudes com 96,4% de precisão.

Este artigo propõe uma metodologia para detectar e classificar ataques de negação de serviço distribuídos. A metodologia proposta emprega técnicas de balanceamento de dados, pré-processamento e seleção de atributos diferentes das encontradas nos trabalhos relacionados. São avaliados cinco algoritmos de aprendizado de máquina e o conjunto de dados usado para treinamento, validação e avaliação é o CIC-DDoS2019. Experimentos mostram que o algoritmo Random Forest (RF) apresenta os melhores resultados tanto na classificação binária quanto na classificação multiclasse. No cenário binário sem dados sintéticos, o RF atingiu 99,8% de acurácia, enquanto na classificação multiclasse alcançou uma taxa de detecção de 100% para ataques SYN e 98% ou superior para outros tipos de ataques.

This work presents an advanced cognitive architecture for networked aerial robots to implement autonomous swarm systems effectively. It focuses on designing, implementing, and evaluating an architecture that enables unmanned aerial vehicles (UAVs) to coordinate and cooperate for complex tasks, with or without human intervention. Inspired by artificial intelligence, cognitive science, and robotics, the architecture integrates perception, planning, decision-making, and adaptive learning to optimize swarm behavior in dynamic environments. The architecture uses a distributed processing model based on the “edge-fog-cloud” (EFC) concept. Edge-level robots handle real-time data collection, local decision-making, and environmental perception. Fog-level vehicles manage intermediate processing and supervision of the groups, while cloud servers perform comprehensive data analysis and long-term storage, being the higher-level hierarchy of the framework. This structure allows efficient distribution of computational tasks, with critical decisions made at the robot level and complex analysis done in the fog or cloud. The implementation of ARCog-NET involves deploying a multi-agent simulation system using the Robot Operating System (ROS) and Gazebo simulator, facilitating the integration of sensors, communication protocols, and data processing algorithms. The performance evaluation demonstrates the architecture’s effectiveness in a wind farm inspection scenario, where the UAV swarm exhibits improved trajectory planning, collision avoidance, and data processing efficiency. Simulation results show that ARCog-NET reduces latency, increases data throughput, and enhances operational effectiveness, providing a robust platform for future developers to focus on applications and direct robot control methods.

Wireless Sensor Networks (WSNs) have become ubiquitous across various domains, from household applications to industrial grounds, and critical areas like emergency response and environmental monitoring. While numerous optimization techniques have emerged in the literature to enhance network performance by strategically allocating sensor nodes, most studies in this field rely on heuristic methods. Designing the topology of WSNs is crucial for optimizing node energy consumption, connectivity, and coverage area. In this work, two methodologies are proposed to minimize the overall energy consumption of the network. The first is an exact method using a Mixed Integer Linear Programming (MILP) model. The second is an optimization method based on genetic algorithms (GAs). In the computational experiments, instances were generated based on grid-like networks, some of which assuming highly irregular topologies. The results have shown that the exact approach found optimal solutions for some instances, up to five thousand nodes. In contrast, the GA obtained high-quality solutions for instances with more than 18 thousand nodes in reasonable computational times.

This paper proposes the Set-Membership sign-NLMS (SM-sign-NLMS) adaptive filter, which combines the ability for data censoring (offered by Set-Membership schemes) with robustness against impulsive noise (provided by signed schemes). The algorithm can present a much lower steady-state probability of update than the standard SM-NLMS algorithm when impulsive noise is present in the system. It is derived from a local deterministic optimization problem modulated by a minimum disturbance cost function combined with a bounded error criterion. Several stochastic models are proposed in order to extract insights and a time-variant step size extension of the algorithm. The first of them, based on energy conservation arguments, leads to a fixed-point analytic equation whose solution predicts the asymptotic performance of the algorithm. Further, a transient analysis based on a statistical decoupling of the radial and (discrete) angular distributions of the input vector is derived. Based on such an analysis, an efficient time-variant step-size version of the algorithm is proposed. Additionally, such an analysis is also utilized to obtain a fixed-point formula whose solution describes the asymptotic performance when the unknown plant that the filter intends to match varies according to a first-order Markovian model. Lastly, a novel stochastic model is advanced for the description of the algorithm learning behavior under a deficient-length scenario for a white input signal, which provides some insights about the asymptotic performance of the algorithm. The findings are confirmed by extensive simulations.

Visualizing biomedical ontologies is an ongoing and highly relevant challenge in dealing with the vast volume of daily data and the miscellaneous use cases of ontologies. BioPortal is the most well-known tool for visualizing biomedical ontologies, bringing together hundreds of ontologies of varying sizes, depths, and complexities. This article aims to establish a methodological triangulation using SIM-SR based on Semiotic Engineering, the WCAG guidelines from the W3C consortium, and ergonomic principles to evaluate the experience of users who utilize screen readers and users who do not use this feature without conducting user experiments. The objective is to investigate and intersect the usability, communicability, and accessibility of BioPortal. Additionally, it seeks to understand how ontologies can be analyzed by as many people as possible, ensuring they are comprehensible and enhancing users’ cognitive understanding of the explored domain.