Omar Andres Carmona Cortes

Purpose: Compact YOLOv11 nano/small variants were trained on public cervical cytology datasets, harmonized via Bethesda labeling for cells and nuclei at high resolution. Models performed joint detection and segmentation, assessed by mAP, recall, and F1-score per category. Real-time use was tested on video from CMOS cameras on a microscope, tracking inference time and frame rate. Methods: Compact YOLOv11 nano and small variants were trained on public cervical cytology images. These datasets were aligned using a unified Bethesda labeling scheme for whole cells and nuclei at high resolution. The models were trained for joint box detection and segmentation, and their results were assessed via mAP, recall, and F1-score for each Bethesda category. The real-time use of the models was tested by deploying them on video streams from low-cost CMOS cameras on a microscope, recording the per-frame inference time and frame rate. Results: The YOLOv11 small model achieved high mAP and recall for both bounding-box and mask predictions across Bethesda classes while maintaining low latency for real-time applications. Inference on microscope video sequences showed frame rates above the 25 fps threshold for real-time performance, even under tough conditions with overlapping cells and complex inflammatory or hemorrhagic backgrounds. Conclusion: These findings suggest that lightweight YOLOv11-based models can offer fast and reliable analysis of conventional Pap smears directly at the microscope, which can support the development of affordable, point-of-care computer-assisted screening tools for cervical cancer, notably in low- and middle-income countries where access to liquid-based cytology and whole-slide scanners is limited.

In the realm of intelligent systems, addressing the NP-hard Permutation Flow-Shop (PFS) problem presents a notable challenge, especially with burgeoning complexities in manufacturing systems. Conventional wisdom relies on metaheuristic techniques to tackle such intricacies, yet the best metaheuristic remains elusive. This paper introduces a new meta-learning framework that leverages pre-trained metaheuristics rooted in Iterated Local Search (ILS) to forecast the most effective metaheuristic for each PFS instance. Notably, our approach pioneers the extraction of novel meta-features from graph representations, discerning the best metaheuristics in 99.7% of cases using a small set of meta-features and baseline classifiers, such as decision trees and support vector machines. Moreover, the interpretability of our meta-model unveils insights into feature impact, showcasing the efficacy of abstract graph computations, such as centrality and connectivity, over bespoke meta-features related to problem size.

Real-time passenger monitoring in resource-constrained ferry terminals demands low-latency deep learning inference under severe occlusion, high crowd density, and variable lighting conditions. This paper presents a structured YOLOv8 development pipeline for edge deployment on single-board computers (SBCs) that addresses these computer vision challenges in public transportation IoT systems. Trained on a custom 2229-image dataset augmented from EMAP surveillance footage, the model achieves 97% person-detection precision in production conditions, even with 5 classes. Comprehensive benchmarking across Raspberry Pi 3/4B and Jetson Nano reveals critical hardware trade-offs: CPU-only SBCs fail real-time requirements, while Jetson Nano delivers 5.3 ms/image inference via Maxwell GPU acceleration. These findings validate state-of-the-art CNN feasibility for edge IoT, quantifying performance-resource boundaries essential for safety-critical transportation infrastructure. The analysis guides SBC selection for scalable, low-power vision systems.

Reinsurance is a common practice in financial engineering that aims to hedge risk and reduce the probability of bankruptcy by obtaining essential services from more prominent insurance companies. Hence, insurance companies might issue contracts in the market to transfer risk, especially against natural disasters. To perform the risk hedge, it is crucial to offer the market the best contract option, i.e., an optimized reinsurance contract, which is a combinatorial problem with a complex search space. In this context, metaheuristics have proven to be practical algorithms for discovering solutions to complex search spaces in a feasible time. Thus, this paper proposes a modified bio-inspired metaheuristic, the Enhanced Multiobjective Particle Swarm Optimization (E-MOPSO), for reinsurance contract optimization, allowing insurers to hedge the maximum possible amount of risk through reinsurance and protect insurance companies against massive claims resulting from natural catastrophes. To verify the effectiveness of those multiobjective solutions, the algorithm was implemented in R and compared against state-of-the-art algorithms such as MOPBIL and VEPBIL, using ANOVA and the Tukey test in a real-world contract comprising 7 layers. Moreover, a 15-layer semi-synthetic dataset was created using the original one to test the algorithm’s ability. Results show that the E-MOPSO approach overcomes the referred algorithms regarding computing time and hypervolume.

Este artigo analisou de que modo perfis de engajamento estudantil, obtidos por clusterização não supervisionada, podem apoiar a identificação de vulnerabilidade acadêmica e a formulação de recomendações contrafactuais interpretáveis para o acompanhamento educacional. O estudo adotou abordagem quantitativa, aplicada e exploratória, utilizando dados secundários do Open University Learning Analytics Dataset (OULAD). A metodologia envolveu pré-processamento, seleção de variáveis comportamentais do ambiente virtual de aprendizagem, clusterização por K-Means, avaliação da qualidade e estabilidade dos agrupamentos e geração de recomendações contrafactuais restritas a variáveis acionáveis. Atributos demográficos e desempenho final foram excluídos da clusterização, sendo reservados à interpretação e validação externa dos perfis. Os resultados evidenciaram um gradiente de engajamento, com perfis de baixa, média e alta participação. Grupos de baixo engajamento concentraram maiores proporções de desistência e reprovação, indicando associação com vulnerabilidade acadêmica. As recomendações contrafactuais permitiram simular transições para perfis de maior participação e mostraram-se viáveis na maior parte das transições analisadas. Conclui-se que a abordagem proposta pode apoiar a tomada de decisão educacional, desde que suas recomendações sejam compreendidas como subsídios estatísticos auditáveis, e não como prescrições automáticas ou inferências causais.

This investigation examines the performance of students from Maranhão, Brazil, in the 2023 edition of the National High School Exam (ENEM) using unsupervised machine learning to uncover latent patterns in large-scale educational data. Leveraging the CRISP-DM framework on microdata from millions of ENEM participants, we applied Recursive Feature Elimination (RFE) with a Random Forest classifier to select key socioeconomic variables, followed by association rule mining using the FP-Growth algorithm across multiple experimental configurations. The results reveal strong associations between low academic performance and factors such as parental education and occupation, lack of household technology (e.g., computers and washing machines), and gender. These findings demonstrate the utility of unsupervised learning for descriptive educational analytics, providing practical insights for targeted policy-making, resource allocation, and regional equity monitoring. This research addresses the underrepresentation of Maranhão in the educational data mining literature and proposes a scalable analytical framework applicable to other developing regions. Despite limitations in data completeness, the approach offers a replicable model for using artificial intelligence to inform public education strategies in socially vulnerable areas.

A escolha de uma implementação eficiente do algoritmo de Dijkstra para encontrar o caminho mais curto em grafos é crucial, especialmente em aplicações que lidam com grandes volumes de dados, como redes de transporte, sistemas de navegação e redes de telecomunicações. Este estudo investiga e compara diferentes abordagens paralelas do algoritmo de Dijkstra, incluindo implementações com OpenMP, MPI e CUDA, em relação à versão sequencial tradicional. As implementações paralelas foram avaliadas quanto ao tempo de execução, speedup e eficiência, utilizando grafos de diferentes tamanhos. Todas as versões paralelas superaram a implementação sequencial em desempenho, com variações de eficiência conforme a abordagem e o tamanho do grafo, alcançando speedup de 6,31 com 8 threads em OpenMP, 6,64 com 4 processos em MPI e 37,8 em GPU para grafos com 16.384 vértices.

O câncer de mama é uma doença responsável pela maioria das mortes no Brasil e afeta principalmente mulheres. Entre as consequências de sua ocorrência estão a predisposição genética, o sedentarismo e a menopausa tardia. É uma doença que requer diagnóstico o mais rápido possível para que o paciente possa iniciar o tratamento. Além disso, estudos sugerem que patologistas podem atingir uma precisão de cerca de 72% ao analisar um exame composto por milhares de imagens histopatológicas de cortes de linfonodos. Nesse contexto, este trabalho descreve um classificador que combina ResNet50, Random Forest (RF) e Support Vector Machine (SVM) treinados pelo conjunto de dados PatchCamelyon. Os resultados indicam que o método proposto atingiu uma precisão e F1-score de 81%.