FCT: DF - Dissertações de Mestrado
URI permanente para esta coleção:
Navegar
Entradas recentes
- Characterization of radiation beams for metrology in Radiological Protection by X-ray spectrometryPublication . Karim, Maria; Kling, Andreas; Cruz, JoãoEnsuring accurate metrological verification of ionizing radiation monitoring equipment is a crucial task in the area of radiological protection of exposed workers and patients subjected to ionizing radiation. This work aims to enhance the metrological verification of X-ray detection equipment at the Portuguese Metrology Laboratory of Ionizing Radiation (Laboratório de Metrologia de Radiações Ionizantes, LMRI), located at the Technological and Nuclear Campus of Instituto Superior Técnico, University of Lisbon. The study explores the application of X-ray spectrometry employing cadmium telluride detectors to perform systematic studies on the properties of the X-ray beams available at LMRI allowing for the extension of calibration conditions beyond those currently available. For this purpose, the energy spectra of the beams were measured as a function of the distance from the source, tube current and acceleration voltage. Monte Carlo methods were employed to assess the detector’s response to monoenergetic radiation and to build a response function matrix for reconstructing continuous X-ray spectra from the measured data using deconvolution algorithms. As an additional asset, methods for determining the spectral endpoint were implemented based on a machine learning algorithm offering an alternative approach to precisely determine the acceleration potential of the X-ray tube.
- Deteção de traços moleculares de triclosan em meios líquidos complexosPublication . Cunha, Gonçalo Teófilo Alves da; Henriques, Célia; Pires, Ana; Raposo, Maria de FátimaO triclosan (TCS) é um agente antibacteriano utilizado em produtos farmacêuticos e de hi- giene pessoal (PPCP), cada vez mais recorrente em reservas de água, o que representa riscos para a saúde pública. Torna-se essencial o desenvolvimento de métodos de deteção de TCS para monitorização. Devido às propriedades lipofílicas do TCS, o objetivo desta tese consis- tiu na criação de um sensor baseado em filmes ultrafinos de lípidos, capaz de detetar quan- tidades vestigiais de TCS em meios aquosos complexos. Para concretizar este objetivo, este trabalho foi dividido em duas partes: Na primeira, estudou-se a interação da molécula de TCS com uma bicamada lipídica de 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC) utilizando simulação de Dinâmica Molecular (MD). Verificou-se a fácil inserção da molécula de TCS na bicamada, preferencialmente entre 9 e 11 Å. O estudo do ângulo de inserção indicou liberdade de rotação, evitando orientações em que os cloros mais periféri- cos se alinhavam em paralelo com a normal da bicamada. Na segunda parte, desenvolve- ram-se sensores baseados em filmes lipídicos e analisaram-se as suas características elétricas em função da concentração de TCS em água da torneira. Suportes sólidos com elétrodos interdigitados de ouro, revestidos por filmes finos preparados por camada-por-camada uti- lizando o polieletrólito hidrocloreto de polialilamina (PAH) e o fosfolípido 1,2-Dipalmitoyl- sn-glycero-3-phosphoglycerol (DPPG), detetaram TCS em amostras de água da torneira do- pada utilizando espectroscopia de impedância como método de deteção. O LOD do sensor foi de 4.23 × 10⁻11 M e detetou-se deriva devido à possível degradação dos filmes. Os dados foram analisados por método de análise de componentes principais (PCA), que demonstra- ram a capacidade de distinguir claramente as amostras com diferentes concentrações de TCS num intervalo linear de 10-12 a 10-5 M. O PCA feito a partir dos valores de ângulo de fase e módulo de impedância, apresentou uma sensibilidade de 3.51 ± 0.21, sendo possível cal- cular a concentração de TCS em função de F1 através da calibração 𝐶 = 10𝐹1/3.51. Estes resultados permitem concluir que filmes lipídicos apresentam elevado potencial para o de- senvolvimento de uma língua eletrónica que permita detetar TCS em água
- Development of a Kinematic Data Analysis Pipeline for Patients with Lumbar Spinal StenosisPublication . Amaral, Afonso Miguel Abrantes de Brito; Mendes, César; Silva, LuísLumbar spinal stenosis (LSS) is a degenerative condition characterized by narrowing of the lumbar spinal canal, leading to pain typically exacerbated by standing and walking. Depending on severity, management ranges from conservative care to surgery. In practice, however, treatment decisions often rely heavily on subjective information. Similar challenges exist across many other pathologies, where objective biomechanical data could improve medical decision-making. Inertial measurement units (IMUs) provide a non-invasive and accessible means to capture patient gait kinematics, enabling more objective assessments. Yet, current gait analysis software is limited, most closed-source solutions output only tabular parameters, restricting transparency, reproducibility, and, critically preventing access to full time-series data required for advanced analyses such as feature engineering for pathology stratification. To address these limitations, the aim of this thesis was to develop and validate an open-source pipeline that transforms raw IMU signals into clinically interpretable gait parameters while preserving access to the underlying time-series. To achieve this, an automated workflow was implemented to process accelerometer, gyroscope, and magnetometer signals from 15 IMUs through preprocessing, sensor fusion, and biomechanical modeling in OpenSim. Joint-angle Validation was performed against an optical motion capture reference. The system successfully generated reliable lower-limb kinematics, with strong agreement against motion capture, while upper-limb kinematics showed inconsistent validity. From validated joints, several features discriminated healthy from patient groups, with large effect sizes in more severe cases. Although discrimination between conservative and surgical patients was limited, the framework proved robust for detecting pathological deviations in gait. This work delivers a transparent and reproducible end-to-end system that enables objective and explainable gait assessments, while allowing advanced feature extraction for pathology stratification. It demonstrates the clinical potential of IMU-based gait analysis to support decision-making, enhance treatment precision, and improve patient outcomes.
- Minimum Viable Dataset: Towards Data-efficient Machine LearningPublication . Capelo, Bárbara Sofia Francisco; Gamboa, Hugo; Folgado, DuarteHealthcare generates petabytes of data each year, creating unique avenues for scientific progress, but also significant computational challenges for training Deep Learning models. Although performance has long followed the “scale-is-everything” paradigm, where larger datasets and models yield better results, emerging evidence shows that carefully curated smaller datasets can achieve comparable performance far more efficiently. This dissertation investigates how to compress large collections of healthcare data into smaller, efficient datasets that preserve task-relevant knowledge while easing com- putational demands. To this end, it introduces the Minimum Viable Dataset (MVD), a minimal yet highly informative subset of real data, optionally enriched with synthetic samples, designed to retain essential knowledge while substantially reducing training requirements. Previous work has focused on developing and benchmarking data summa- rization methods using performance metrics alone, which often overlook dataset diversity and representativeness. This dissertation takes a complementary approach, using data quality metrics to guide both the creation and evaluation of summarization methods, providing a more principled path toward efficient and reliable model training. The main contributions are threefold: (1) the formalization of the Minimum Viable Dataset (MVD) concept and its taxonomy; (2) a systematic study of data quality metrics for evaluating subsets; and (3) a hybrid strategy that leverages coreset-informed subsets to initialize dataset distillation methods. These approaches were primarily developed and validated on a dataset of hematological cell images, and further assessed across additional standardized biomedical image collections. The results show that MVDs provide a reproducible framework for efficient dataset construction and yield design directions for future dataset summarization methods. By retaining essential knowledge while reducing training requirements, MVDs are positioned as a step toward scalable and resource-conscious AI in healthcare, with particular relevance for low- and middle-income countries where computational resources are limited.
- Study of phase modulation techniques to concentrate laser radiation in turbid media for phototherapyPublication . Cruz, Ricardo Manuel Ramos; Coelho, João; Vieira, PedroDiagnostic and therapeutic techniques leveraging electromagnetic radiation as their pri- mary effector are abundant in the context of biomedical engineering. Its interaction mecha- nisms with matter though are often difficult to control, with this unpredictability stemming mainly from scattering phenomena for biological tissues. This study aims to address this through iterative feedback-based phase-only wavefront modulation. As such, an experimental testing procedure where two different feedback mech- anisms are evaluated is implemented. For both setups an 808 nm wavelength laser beam is used, with its already lower absorption and scattering characteristics for biological tissues mak- ing it relevant for potential applications. Also common to both experimental setups are the two different tissue simulating optical phantoms considered. Set to mimic the optical proper- ties of 1.1 mm thick human skin, they were composed of a mixture of agar, distilled water and milk, with the concentration of milk setting the scattering behavior of the turbid samples. The first feedback signal to be tested was gathered by a NIR camera placed behind the phantoms. The latter received the transmitted laser beam's profile after it had been modulated by a spatial light modulator (SLM). This feedback is then sent to the computer and processed in MATLAB by the iterative optimization algorithms implemented. These will finally send new phase masks to the SLM and continue this feedback loop until a desired number of iterations is reached. Given the placing of a sensor behind the tissue is not feasible for in-vivo applica- tions, in view of making the present study hold its relevance for such frameworks, the use of backscattered radiation as feedback to the modulation procedure was scrutinized. For both feedback mechanisms and scattering environments the population-based op- timization algorithm largely provided the best results in comparison with its single solution counterparts. No discernable differences were reported between different turbidity levels, proving the reliability of both techniques across the reported human skin scattering range. The backscattered feedback, although having shown to provide a noisier setting, with this remain- ing largely unaddressed given it was inherent to the available material, it nevertheless proved to be a valid framework for electromagnetic wave propagation management in the absorption and scattering regimes found in biological tissue.
- Concept Learning: Understanding and Extracting Domain-specific concepts in medical ObservationsPublication . Silva, Matilde Gonçalinho; Gamboa, Hugo; Carreiro, AndréIn clinical environments, high-risk decisions require trust and understanding, which can hinder the use of high-performance models such as deep neural networks due to their complexity and opacity. Despite the development of numerous techniques aimed at enhancing interpretability and transparency, the explanations provided are not always human-interpretable and remain underexplored as a means of improving downstream tasks. Furthermore, clinical data often faces the challenge of limited annotated data, despite a large volume of unlabeled data. This study aims to extract human-interpretable concepts from X-ray images and respective reports using multimodal learning with self-supervision, which allows learning directly from the input data without annotations. Moreover, the objective is to enhance the interpretability and performance in clinical tasks. Therefore, the work integrates a vision language model with self-supervision techniques to extract domain-specific textual concepts from the joint representation of X-ray images and reports, while also investigating changes in the model framework to improve the alignment between the concepts and image features. Finally, these extracted concepts are incorporated as features into an X-ray image classification model to assess their impact on performance. A key contribution is the development of a novel dataset generator with controllable ground truth concepts, providing a flexible and scalable tool to test concept-based ap- proaches and identifying appropriate methods for medical data. Additionally, the results demonstrate that the incorporation of large language models within self-supervised vision-language frameworks significantly enhances the extraction of concepts, leading to direct improvements in diagnostic classification performance and interpretability. These findings highlight the potential benefits of integrating multimodal frameworks with self-supervision for concept extraction, thereby establishing a foundation for more inter- pretable, transparent, and reliable artificial intelligence-driven healthcare solutions.
- O papel de eletrões de baixa energia em fotorresistes de estanho para EUVL: Identificação e análise de espécies iónicas formadas a partir de resistes moleculares organometálicosPublication . Sequeira, Diogo Arroja; Silva, Filipe; Mendes, MónicaEm litografia, o comprimento de onda da luz utilizada para gravar padrões num fotorresiste está fortemente relacionado com o tamanho mínimo das estruturas que são possíveis de obter. Nas últimas décadas, tem sido utilizada litografia com luz na gama do ultravioleta profundo (DUVL), capaz de atingir resoluções de cerca de 40 nm, onde por resolução se entende a menor dimensão geométrica que pode ser fielmente reproduzida num fotorresiste. No entanto, a litografia com luz na gama do extremo ultravioleta (EUVL), que utiliza um comprimento de onda de 13,5 nm (92 eV), tem sido introduzida como a tecnologia predominante no fabrico de semicondutores. Um dos principais obstáculos à transição entre DUVL e EUVL reside na necessidade de desenvolver fotorresistes compatíveis com este novo processo. Os fotorresistes am- plificados quimicamente (CARs) mostram ser uma solução viável para DUVL, porém, a transferência para EUVL tem sido problemática devido à radiação altamente energética desta nova técnica e à baixa secção eficaz de absorção de EUV. Face a estes obstáculos, a inclusão de metais em novos sistemas de fotorresistes tem sido um dos focos primários dos fabricantes. Está estudado que os eletrões secundários de baixa energia gerados a partir da fotoio- nização das moléculas do filme são responsáveis pela maioria das reações no fotorresiste. Com este fenómeno em mente, esta dissertação de mestrado procura estudar as interações de dois agregados oxo-metálicos diferentes com eletrões de baixa energia, bem como os fragmentos iónicos provenientes de tal reação, de modo a expandir o conhecimento existente sobre sistemas de fotorresistes viáveis para a realização de EUVL. Experimentalmente, foi utilizado um espetrómetro de massa de tempo-de-voo com refletrão ortogonal, em combinação com um monocromador trocoidal de eletrões, destinado a colimar o feixe de eletrões que intersecta o feixe molecular. Para além da identificação dos fragmentos iónicos através da obtenção de um espetro de massa das moléculas, serão calculadas as energias de aparecimento, para catiões, e ressonâncias, para aniões.
- Comparative Analysis of Recurrent Neural Networks Performance for the Optimization of a Gamified Neurofeedback Brain Computer Interface in Anxiety TherapyPublication . Barbosa, Miguel Ângelo Louro; Belo, David; Ferreira, Hugo; Vigário, RicardoAccording to its 2025’s report, the World Health Organization (WHO) identified anxiety disorders as the most common mental health condition, affecting more than 359 million indi-viduals, from which less than 25% receive treatment which is many times costly or inefficient. Once modern psychiatry could benefit from more innovation in these field, this research aimed to personalize an existing gamified neurofeedback solution for anxiety control therapy using less computational expensive AI models. To this end, a dataset containing time-series of EEG derived data, from 19 patients playing in a gamified neurofeedback BCI. Each game’s full session was divided in 10 subsessions of data, each corresponding to a game’s level. From each game’s level 82 features were extracted but only the power average asymmetry of alpha waves in channels F3-F4 (APAA F3-F4) was used, in accordance with prior research identifying it as the best metric for accessing anxiety control for this solution. This research developed and tested different architectures of LSTM and GRU models, with different sets of hyperparame-ters while proposing a new pipeline of innovation in neurofeedback which can improve future lines of research. Results showed that algorithms show potential to predict anxiety alterations using APAA F3-F4 in real world applications. Several architectures achieved a mean direc-tional accuracy (MDA) above 0.6 for 8-seconds windows, representing 6 times better results than previous models. Root mean squared error (RMSE) was kept under 20% for multiple models. Furthermore, correlations were observed between game level and model performance indicating that space for improvement.
- Development and Implementation of Machine Learning- and Physics-Based Models for Wildfire PropagationPublication . Marujo, Gonçalo da Silva; Martins, Leonardo; Vieira, PedroThe New Space Portugal project is a consortium of 39 entities aiming to design, develop, produce, and launch Earth Observation (EO) satellites, positioning our country at the forefront of global markets in this sector. More specifically, Bee2Fire, in collaboration with NOVA School of Science and Technology, is part of Work Package 4 (Digital Planet) and contributes to the development of algorithms for modeling the propagation of wildfires, based on data acquired by satellite image sensors. Wildfires can have devastating consequences, frequently causing human casualties. They can destroy vast forests, homes, and habitats, leading to the displacement of animals and loss of biodiversity. Moreover, the financial impact can be significant, as firefighting efforts are costly, and recovery from the damage caused takes years, affecting local economies and infrastructure. Within this context, the present dissertation focused on the development and implementation of wildfire propagation models based on machine learning and/or physical principles, relying mainly on the Python programming language. In particular, two approaches were explored: Next Day Wildfire Spread, a deep learning model trained on historical wildfire data from the United States; and ForeFire, an open-source wildfire simulation engine that supports a variety of spread models. A set of real occurrences with diverse characteristics was selected to evaluate both approaches. Each simulation was validated against available ground truth data and subsequently compared to results obtained with FARSITE, a widely used benchmark tool. Predictions were analyzed and discussed, highlighting the strengths and limitations of each approach. While both demonstrated potential, differences in performance were observed depending on the nature of the event: Next Day Wildfire Spread is ideal for extreme weather conditions, while ForeFire is more effective in wind-driven scenarios. The study contributes to the ongoing effort to enhance wildfire forecasting capabilities and offers insights for future developments.
- SUSTAINABLE SOFT SKIN BIOELECTRONICS FOR CYSTIC FIBROSIS DIAGNOSISPublication . Franco, Pedro Miguel da Silva; Fortunato, Elvira; Pinheiro, TomásCystic Fibrosis is a chronic hereditary genetic disease that significantly reduces the life expectancy of the carrier, causing various health problems that tend to worsen over the years. Therefore, it is important to make a quick diagnosis to begin the recommended treatments to allow for a longer life expectancy and better quality of life for the patient. The most common diagnostic tests are performed weeks after birth by collecting sweat using iontophoresis tech-nique, where chloride ion levels are subsequently analyzed. High values (≥60 mM) are indic-ative of the disease, although values between 30 and 59 mM may also be a sign of it, requiring further testing. However, traditional equipment is expensive, complex, and requires skilled labor to perform the tests, making diagnosis difficult in less developed countries. The objective of this dissertation project was to develop a low-cost, easy-to-use colori-metric device using environmentally friendly and sustainable materials to enable early diag-nosis of Cystic Fibrosis. This device was based on laser-induced graphene (LIG) produced from cellulose-based materials using the iontophoresis technique. This technique allows sweat to be produced using a pilocarpine hydrogel, which penetrates the skin through the passage of electric current. To enable the diagnosis, the reagent silver chloranilate is used, which dis-plays different shades of purple depending on the concentration of chloride ions in the sweat. The device showed very positive results, with chloride values of 37.7 ± 3.1 mM. In ad-dition, the test time was significantly reduced (35 minutes) and a low electric current density (0.33 mA.cm-2) was used, ensuring maximum user comfort.