https://tedebc.ufma.br/jspui/handle/tede/1314 Dissertações defendidas no âmbito do Programa. 2026-07-04T17:12:37Z https://tedebc.ufma.br/jspui/handle/tede/7059 Título: Método não supervisionado de sumarização extrativa de textos jurídicos com alinhamento de grafos semânticos guiados por atenção Autor: BERNHARD, Pedro Vinnícius Primeiro orientador: ALMEIDA, João Dallyson Sousa de Abstract: The massive volume and technical complexity of legal documents in Brazil impose a major challenge to the efficiency of the judicial system. Automatic summarization emerges as an alternative to mitigate this overload and assist the work of judges and lawyers. However, the application of deep learning models in Law faces critical obstacles: algorithmic opacity (“black-box”), the unacceptable risk of factual hallucinations in generative models, and the severe scarcity of labeled data for training. Thus, the development of solutions that unite factual fidelity and interpretability is essential. In this context, this work proposes an unsupervised extractive summarization method focused on the legal domain, structured on the modeling of semantic graphs guided by attention mechanisms. The method extracts self-attention weights from an expert language model (Legal-BERTimbau) and filters noisy connections via dynamic binarization using Otsu’s method. The text is converted into a directed graph, thematically partitioned by the Hierarchical Infomap algorithm to isolate the argumentative axes. Topic alignment is performed in a dense vector space (Sentence-BERT), and sentences are ranked by the Maximum Attention heuristic, respecting a strict compression limit of 10%. In the evaluation using the RulingBR dataset, the proposed model outperformed classical unsupervised algorithms in the ROUGE-1 (36.61%) and ROUGE-L (20.74%) metrics. Additional experiments with an Extractive Oracle demarcated the upper bound of the task at a ROUGE-1 of 65.21% and ROUGE-L of 47.37%, while a hybrid extractive approach guided by an LLM (GPT-5 mini) achieved a ROUGE-L of 21.31%. Thus, the developed method proves promising by ensuring the integrity of the original text, free from procedural hallucinations, additionally offering a visual explainability interface that makes sentence selection fully auditable. Instituição: Universidade Federal do Maranhão Tipo do documento: Dissertação 2026-04-29T00:00:00Z https://tedebc.ufma.br/jspui/handle/tede/7027 Título: Atenção Multiescala em Redes U-Net: Uma Abordagem para Segmentação de Rins, Tumores e Cistos em Tomografia Computadorizada Autor: OLIVEIRA, Marcus Vinicius Silva Lima de Primeiro orientador: BRAZ JUNIOR, Geraldo Abstract: Early detection and diagnosis of renal cancer play a crucial role in patient prognosis and treatment, significantly increasing the chances of survival and cure. In this context, advances in radiological imaging have enabled medical specialists to analyze and identify suspicious lesions more effectively. Computed Tomography (CT) stands out as a widely used tool for renal cancer diagnosis due to its ability to generate high-resolution images of internal body structures, including the kidneys, cysts, and tumors. However, manual analysis of these exams is time-consuming and error-prone, often affected by fatigue and distraction, highlighting the need for computational methods to support the automatic segmentation of these structures. In recent years, Convolutional Neural Networks (CNNs) have shown significant potential in this task, providing valuable support to medical imaging specialists. In this study, we propose a convolutional model named Dual-Scale SE U-Net, which leverages multi-scale feature extraction through parallel convolutions (3×3 and 7×7), combined with channel attention mechanisms based on the Squeeze-and-Excitation (SE) module, integrated into the U-Net architecture for the segmentation of kidneys, cysts, and tumors in CT images. As a preprocessing step, the images were resized from 512×512 to 256×256 pixels to improve computational efficiency. The proposed methodology achieved promising results on the KiTS23 dataset, evaluated using five-fold cross-validation, yielding Dice similarity coefficients of 90.94% for Kidneys and Masses, 89.52% for Renal Masses, and 86.27% for Renal Tumors. In the analysis of individual structures, the model achieved 93.80% for kidneys, 92.76% for cysts, and 86.27% for tumors, demonstrating the effectiveness of the proposed approach as a supportive tool for medical diagnosis. Instituição: Universidade Federal do Maranhão Tipo do documento: Dissertação 2026-04-23T00:00:00Z https://tedebc.ufma.br/jspui/handle/tede/7024 Título: Um simulador de trajetória de VANT de asa fixa aplicado ao monitoramento espectral de radiofrequência Autor: MARCOS, Paulo Roberto Silva Primeiro orientador: OLIVEIRA, Alexandre César Muniz de Abstract: The use of unmanned aerial vehicles (VANTs) in radio frequency spectrum monitoring activities is one of the many possible applications of this type of vehicle, whether operating autonomously, pre-programmed to follow a planned route, or remotely controlled, and several studies have proposed algorithms for this purpose in different application contexts. This research proposes a trajectory simulator to be executed by a fixed-wing VANT, suggesting routes that include horizontal and vertical deviations during spectral monitoring of radio frequency signals emitted by known transmission stations, while also assisting in the identification of signals from unknown stations along the path. As part of the planning stage, the approach is based on refining sequences generated by basic routing algorithms used in the traveling salesman problem (TSP), as well as through modeling in the Gekko optimization suite and APOPT solver, combined with the definition of the search space derived from radio frequency characteristics of real stations and signal-to-noise ratio thresholds calculated using a propagation model appropriate to the monitored band and the VANT’s hypothetical flight altitude. The results demonstrated that the minimization achieved by the hill-climbing algorithm led to a significant reduction in trajectory length, which was adequately simulated by the proposed model. The simulation considered the presence of large obstacles requiring altitude variation, as well as horizontal deviations to bypass smaller obstacles, aided by the energy consumption function, allowing prediction of return moments for refueling and continuation of the initially proposed sequence. Instituição: Universidade Federal do Maranhão Tipo do documento: Dissertação 2026-04-15T00:00:00Z https://tedebc.ufma.br/jspui/handle/tede/7009 Título: Diagnóstico automático da paralisia do sexto nervo em vídeos de motilidade ocular por classificação de séries temporais e aprendizado profundo Autor: FERNANDES, Saulo Enock Rodrigues Primeiro orientador: ALMEIDA, João Dallyson Sousa de Abstract: The sixth cranial nerve innervates the lateral rectus muscle, which is responsible for left-to-right eye movements. Paralysis of this nerve impairs the lateral rectus muscle’s function, potentially causing headaches, migraines, blurred vision, dizziness, and diplopia (double vision) when attempting to move the eye toward the outer corner. Therefore, early diagnosis of this condition is essential to prevent long-term effects. Since available diagnostic techniques are invasive or expensive, this study proposes an automated method for diagnosing sixth cranial nerve palsy based on classifying time series extracted from eye-tracking data of ophthalmological patients using deep learning models, aiming to assist specialists in the diagnostic process. The proposed method uses the YOLO network to track eye movement accurately and quickly in ophthalmological videos, along with the MediaPipe facial detection model, in a strategy to compensate for eye trajectories, removing occasional head movements and isolating the actual eye movement. In a cross validation experiment, the proposed method achieved a mean sensitivity of 70% across the folds and 75% in the best-trained fold, indicating that our methodology has potential for clinical application. Instituição: Universidade Federal do Maranhão Tipo do documento: Dissertação 2026-05-04T00:00:00Z