FCT: DF - Teses de Doutoramento
URI permanente para esta coleção:
Navegar
Entradas recentes
- Machine Learning–Based Characterization of Adrenal Adenomas in MR Images Incorporating Deep Learning Detection, Radiomics, and GAN-Based Image SynthesisPublication . Gonçalves, Bernardo Brás; Vieira, Pedro; Vieira, Ana LuísaAdrenal lesions are relatively common and are often detected incidentally in abdomi- nal magnetic resonance imaging (MRI) exams. Correct characterization of these lesions, namely the distinction between benign adenomas and malignant tumors such as metas- tases or carcinomas, is crucial to avoid unnecessary interventions and ensure appropriate treatment. However, conventional image-based diagnosis using MRI is a complex and time-consuming process, largely dependent on the radiologist’s experience. This thesis proposes a hybrid machine learning pipeline aimed at automating the detection and classification of adrenal lesions in multi-sequence MRI images. The system comprises two main stages: in the first stage, a detection model based on the FCOS architecture automatically identifies lesion regions; in the second stage, an ensemble classifier, based on radiomic features extracted from the lesions, distinguishes between adenomas and non-adenomas. The pipeline was developed and validated using a dataset collected from Garcia de Orta Hospital, which was carefully pre-processed and annotated with the support of medical specialists. In parallel, a generative model based on the StyleGAN3 architecture was also developed to synthesize realistic abdominal MRI images, with the aim of mitigating data scarcity and class imbalance. The synthetic images were evaluated quantitatively (FID = 12.89; KID = 7.06 × 10−3) and qualitatively by a medical expert, who identified 81% of the generated images as real. However, these images were not integrated into the training pipeline due to the absence of lesion-level annotations. The complete pipeline, requiring no manual segmentation, achieved promising results: an accuracy of 87.45%, a sensitivity of 87.63%, and a specificity of 87.33% in the detection of adenomas. These results demonstrate the viability of the proposed approach as a clinical decision-support tool for the diagnosis of adrenal lesions, with potential for integration into radiological workflows.
- Advances in Multirod Solar-Pumped Laser TechnologyPublication . Costa, Hugo Filipe Cabrita; Liang, Dawei; Almeida, JoanaSolar-pumped lasers are perceived as an innovative convergence of renewable energy and laser technology. Having a long-standing history, they have been the focus of continuous research and development by various teams worldwide since the 1960s. This sustained effort has led to the recognition of solar-pumped lasers as a promising technology for the future, capable of delivering laser radiation in a cost-effective way. In recent years, there has been a rising interest in employing multiple active media to address the thermal load challenges commonly associated with solar-pumped lasers. This thesis explores this approach through a series of numerical works conducted using Zemax® and LASCADTM software, showing that it holds significant potential to improve efficiency, beam quality, scalability, and stability of the laser output power. Furthermore, and most importantly, the results of two experimental studies carried out with a four-rod solar laser prototype are presented. The first work achieved the first-ever simultaneous emission of multiple multimode solar laser beams using a Fresnel lens as the primary concentrator. A total laser power of 22.46W was produced, corresponding to a conversion efficiency of 4.49%. This set a new record for Fresnel lens systems, representing a 1.16-fold improvement over the previous record. Moreover, despite parabolic mirrors having yielded superior outcomes as primary concentrators, the findings indicated that multirod systems incorporating Fresnel lenses can maintain competitive performance while being economically viable. The second work produced multiple TEM00-mode solar laser beams simultaneously for the first time, with a parabolic mirror as the primary concentrator. A total TEM00-mode laser power of 6.05W was achieved, corresponding to a conversion efficiency of 0.90%. This outcome established a new benchmark for solar laser systems incorporating parabolic mirrors, exceeding the previous record by a factor of 1.43.
- Scapular-focused exercise protocol for patients with shoulder dysfunctionsPublication . Santos, Cristina Maria Lopes dos; Gamboa, Hugo; Matias, RicardoBackground: Shoulder dysfunctions, particularly rotator cuff-related pain syndrome (RCS) and anterior glenohumeral instability (ASI), are associated with pain, reduced function, altered scapular neuromuscular control, and impaired movement patterns. Although scapu- lar-focused exercise has been advocated, the evidence supporting its effectiveness, particularly when complemented by electromyographic biofeedback (EMGBF), remains inconsistent. This thesis explored whether a scapular-focused exercise protocol, with or without EMGBF, can improve pain and function, but also scapular neuromuscular activity and control (SNAC), range of motion (ROM) and glenohumeral flexor and abductor muscle strength (GMS) in pa- tients with shoulder dysfunctions, both in the short and long term. Objectives: This doctoral research addressed two main questions: (1) What changes in pain and functional outcomes are observed in patients with shoulder dysfunctions following a scapular-focused exercise protocol supported by real-time EMGBF, and are these outcomes maintained over the long term? (2) In patients with RCS, is a scapular-focused exercise proto- col, with and without EMGBF, superior to a control therapy intervention in terms of pain and function? Methods: Two studies were conducted. Study 1 (Prospective cohort) investigated short- term and 2-year outcomes of a scapular-focused exercise protocol with EMGBF in patients with RCS and ASI. Pain and function (Shoulder Pain and Disability Index - SPADI, Disabilities of the Arm and Shoulder Hand - DASH, Numeric Pain Rating Scale - NPRS), scapular stabi- lizer activation onset (SSAO), scapular stabilizer neuromuscular control (SSNC), dynamic scapular alignment, ROM and GMS were assessed at baseline, 4 weeks, and 2 years follow-up. Study 2 (Randomised clinical trial) compared short-term outcomes of three groups of RCS patients: (a) scapular-focused exercise protocol group (P_G), (b) scapular-focused exercise pro- tocol with EMGBF (P+EMGBF_G), and (c) control therapy group (CT_G) (manual therapy + general strengthening). Assessments were performed at baseline and after 6 weeks. Results: In Study 1, the EMGBF-assisted protocol significantly improved pain, function, SSNC, SSAO, dynamic scapular alignment, ROM, and GMS at 4 weeks. Long-term results showed maintenance of improvements in pain and function, SSAO, ROM, and GMS, although SSNC and dynamic alignment partially regressed over time. In Study 2, all groups improved significantly at 6 weeks. The P+EMGBF_G showed superior short-term improvements in pain and function, SSNC, and dynamic scapular alignment compared with the CT_G, and superior SSNC outcomes compared with the P_G. The P_G showed better pain and function, and dy- namic scapular alignment outcomes than the CT_G. However, pain and function improve- ments were similar between the P_G and P+EMGBF_G, suggesting that EMGBF selectively enhances SNAC outcomes. Conclusions: This thesis provides evidence that a structured scapular-focused exercise protocol is an effective physiotherapy intervention for patients with shoulder dysfunctions, with clinically meaningful improvements in pain and function, but also in SNAC, ROM and GMS. EMGBF offers added value, particularly in enhancing SSNC and dynamic scapular alignment. The long-term findings support the sustainability of most clinical gains, reinforcing the relevance of scapular-focused exercise protocol in physiotherapy practice.
- Learning from Biosignals. A Deep Neural Approach to Advanced Signal ProcessingPublication . Dias, Mariana de Avelino Geraldo; Gamboa, HugoWhen it comes to biosignal processing, a practitioner must undergo a series of steps to prepare the data for analysis. These foundational tasks include assessing data integrity, executing cleansing protocols, standardizing measurements through normalization and unit conversion, and isolating features for the intended analysis. While there are pipelines that guide these procedures, these need to be customized in order to not only adjust to the different characteristics of each dataset but also to the information that is to be extracted. The customization relies on the practitioner’s know-how, underscoring the importance of expertise in transforming raw biosignals into meaningful conclusions. Meanwhile, Deep Learning (DL) frameworks for time-series have consistently demon- strated great potential in performing tasks such as classification and processing of biosig- nals. For that reason, the present work focuses on the application of DL methodologies for biosignals. We start by addressing specific tasks such as disease classification and noise removal in ECG signals. Various strategies are explored, including alternative data representations and architectural paradigms, with a focus on efficiency through reduced model complexity. Additionally, a custom data collection was conducted in industrial environments, and the resulting dataset was used to evaluate the generalization capacity of the noise removal model across different acquisition conditions. Given that a robust DL model needs to learn multiple signal features to, for instance, distinguish between normal and pathological states, leveraging this acquired knowledge across various processing tasks could potentially revolutionize our approach in biosignal analysis, making the case for a more integrated and versatile deployment of DL techniques. In this context, we developed NeuralLib, a framework that provides functionalities to train, reuse and access networks designed to perform biosignal processing functions, built on three key principles: modularity, efficiency, and generalization. These networks learn the fundamental characteristics of the signals, and by fine-tuning them for new tasks, it becomes possible to transfer prior knowledge, enabling not only faster convergence but also reducing the need for large annotated datasets.
- EMI Risk Assessment in Medical Device Innovation Process Towards a Risk-based ApproachPublication . Nunes, Tiago Pinto; Gamboa, Hugo; Silva, HugoRecent reports from American hospitals indicate that the number of incidents caused by Electromagnetic Interference (EMI) in medical devices has been increasing and, with an ever increasingly harsh electromagnetic environment, this trend is expected to remain in a near future. Nowadays, medical devices are no longer circumscribed to clinical environments. From sphygmomanometers to glucose monitors, medical devices have become more and more accessible to the masses and are being used in uncontrolled environments, where the potential threats to the well functioning of these devices is also increasing. The ability of a device to co-habit the same environment as other devices without losses in performance nor disturbing other devices is the definition of Electromagnetic Compatibility (EMC). However, to ensure a device is capable of doing so is progressively harder. Currently, to ensure a device is electromagnetic compatible, standardised tests exist that allow to verify general safety requirements. However, the electromagnetic conditions in which a device will operate are often much more complex (and increasingly so) than those created in common testing settings: these may no longer suffice to declare a device electromagnetically compliant. Thus, a new approach is required to address the issue at hands, an approach that does not focus on pre-determined values only, but rather on settings representative of the real use cases conditions. This risk-based approach calls for a thorough risk analysis of the device considering all the foreseeable scenarios in which it might be placed so that the testing conditions can be set accordingly. In this work, a wearable device intended to be used in home-care scenarios was developed as use case for testing new methodologies to assess and implement a risk-based approach to EMC. This device, dedicated for Electrocardiogram (ECG) acquisition and breathing monitoring, used novel electrodes printed in different materials (silver, carbon and graphene) on a thin substrate to be worn as a patch. A special focus was given on the characterization of these electrodes with particular interest in its behaviour in harsh electromagnetic scenarios. By exposing them individually to a controlled electromagnetic source, it is shown how silver electrodes are more susceptible to environmental noise than the remaining two materials.In addition, the electronics that would serve as reference for the biosignal acquisition system was tested against its electromagnetic susceptibility using a TEM cell, an anechoic chamber and a reverberation chamber. The set of test scenarios simulated in these structures included the use of a skin mimicking phantom to simulate the presence of the human body, demonstrating the effects of the latter in the overall electromagnetic susceptibility of a biosignals acquisition device. Finally, using equipment commonly found in regular laboratories, tests similar to those conducted in anechoic environment were repeated leading to the identical outcomes on the device behaviour and thus allowing to reach identical conclusions. Although they do not aim to replace dedicated test facilities, it is sufficient to demonstrate how simpler solutions allow to reach important conclusions prior to conducting an expensive test campaign in an accredited testing facility.
- Precision Benchmarking of Atomic Data of Highly Charged IonsPublication . Grilo, Filipe Ventura; Amaro, Pedro; López-Urrutia, JoséHighly Charged Ions (HCI) play an essential role in astrophysics, plasma physics, and fusion research, as their emissions are used as key diagnostic tools for evaluating the physical conditions of a plasma. This thesis is devoted to advance and refine the atomic data of HCI for these interdisciplinary areas. From the experimental side, an Electron Beam Ion Trap (EBIT) was used to produce, confine, and probe several charge states of different HCI specimens, observing their emissions in the Extreme-Ultraviolet (EUV) and X-Ray energy ranges. All the observations benchmark previously non-measured theoretical data and new calculations developed on the theoretical side of this thesis. Oxygen is a ubiquitous element in astrophysical observations. Its K𝛼 emissions near the collisional excitation threshold, as well as the Dielectronic Recombination (DR) KL𝑛 satellite structure of its He-like charge state, were measured for the first time. A time- dependent Collisional Radiative Model (CRM) was developed to model the experimental observations. This model was shown to correctly describe the dynamics of the 1𝑠2𝑠 3S1 metastable state, and its results were used to extract the 𝑧-line emissions from the total K𝛼 emission. The dielectronic recombination and excitation structures near the excitation threshold were benchmarked against both literature and new state-of-the-art configuration- interaction calculations. A similar setup was also employed in He-like calcium. For this case, both the cross-sections of the collisional processes and the line positions of the DR satellites were tabulated. The latter could be essential for future high-resolution X-Ray space missions, as He-like calcium lines can be used as an alternative to Ne-like iron lines. Emissions lines of Ne-like krypton for both DR and excitation structures were also studied and benchmarked. The Non-Negative Matrix Factorization (NNMF) technique was shown to be viable for decomposing unresolved emission structures. Moreover, the 𝑛 = 5 intrashell emissions of open 4 𝑓 -shell ions of tungsten were also studied. Observations were made for W10−27+. The NNMF technique was used to obtain individual spectra of every charge state, and the respective emissions were tabulated, and compared with new calculations. Emissions of these elements are vital in the development of X-ray diagnostics for tokamak plasmas, as both krypton and tungsten constitute contamination with charge state distributions given by the plasma local conditions.
- Improving the Robustness of Multimodal AI with Asynchronous and Missing InputsPublication . Santos, Ricardo Bruno Barbeiro dos; Gamboa, Hugo; Carreiro, AndréIntegrating Artificial Intelligence (AI) in clinical Decision Support Systems (DSS) can significantly transform healthcare in many ways to improve patient outcomes. Despite these promises, AI-based clinical DSS face critical challenges, particularly in handling complex, asynchronous, and incomplete medical data. This lack of robustness in AI models and issues around insufficient validation and clinician trust pose substantial barriers to effectively deploying these systems in real-world settings. This thesis addresses these challenges by proposing a holistic framework to develop robust AI-based clinical DSS. The framework is structured around four core strategies: human-centric, data-centric, model-centric, and deployment-centric approaches. Through this framework, we aim to overcome the limitations of current systems and create more valuable and generalizable AI solutions in healthcare. To demonstrate the value of this framework, we explored three medical use cases, each presenting distinct challenges and technical objectives. First, we developed risk models for predicting postoperative complications using a clinician-in-the-loop approach in cardiothoracic surgery. This work focused on addressing real-world data heterogeneity and missing inputs to improve the quality and relevance of the predictions throughout the patient journey. Second, we explored multimodal data integration within cardiovascular disease classification from electrocardiogram signals. We investigated how different data representations in a multimodal framework can be leveraged to enhance model robustness and identified that multimodal schemes are only sometimes superior. Finally, we researched the challenge of handling numerous asynchronous medical variables in continuous monitoring within critical care. We developed ICU-BERT, a transformer-based model designed to effectively process complex, time-sensitive data, improving real-time decision support. To show its generalization potential, we tested ICU-BERT in various databases and real-world tasks. This thesis’s contributions extend beyond technical advancements; they include strate- gies to ensure that AI-based clinical DSS can be integrated effectively into clinical practice. By addressing key challenges in AI model robustness, this work lays the foundation for developing trustworthy clinical DSS that can enhance patient care.
- Glaucoma Detection Using Pupillary ResponsePublication . Sousa, Ana Isabel Araújo Manuel Machado de; Vieira, Pedro; Neves, CarlosGlaucoma is one of the main causes of avoidable blindness. It is a neurodegenerative disease that results in damage in the optic nerve, leading to blindness by the destruction of the gan- glion cells, progressively and irreversibly. An early diagnosis is essential to prevent this type of damage and increase the quality of life of patients. Recent studies show that patients with glau- coma have a diminished pupillary response to blue-light stimuli, with a reduction of intrinsically photosensitive retinal ganglion cells (ipRGCs). Pupillometry, a technique that allows the assessment of the pupillary response to light, has gained renewed interest with the discovery of melanopsin, present in ipRGCs, and its sensitivity to blue light. These cells contribute to the pupillary light reflex, along with cones and rods. Chromatic pupillometry quantifies the pupillary response to blue and red light stimuli. In healthy people, pupil recovery after the blue stimulus is inferior to that of the red stimulus, due to the activation of these photosensitive ganglion cells. The aim of this project was to develop a portable and accessible device to analyse the pupillary response to coloured light stimuli using a smartphone. Initially, the smartphone flashlight was used as a stimulus, but it was not enough to stimulate the ipRGCs. So, after some experiments, a system was developed using a smartphone that allows stimulating the pupil using a coloured ring light. Validated in healthy people, the results were in agreement with those in the literature, with a slower pupil recovery after stimulus for blue light. It has also been evaluated in people with glaucoma. The system developed in this Ph.D. project showed to be a promising portable device for chromatic pupillometry, with stimuli targeted to the activation of intrinsically photosensitive retinal ganglion cells, with the potential to be used for glaucoma screening.
- Electron-Molecule Reactions Application of Computational Methods to Radiosensitizers, Surfaces and Small MoleculesPublication . Romero, José Antonio Rodrigues; Probst, Michael; Limão-Vieira, PauloElectron-molecule reactions cover a very wide variety of interactions such as excitation, ionization, formation/cleavage of chemical bonds, etc. As such, this type of reactions is important in scientific fields ranging from biology and material science to chemistry and physics. From a fundamental point of view, the first step to study these interactions is to obtain the wave function of the participating molecule(s) by solving Schrödinger’s equation since it encodes all the information of the system. However, solving this equation is neither possible analytically for systems with more than one or two electrons without simplifications. Even in the case of numerical solutions the computational cost of solving it increases exponentially with the number of electrons. This is the main concern of ab initio calculations where the wave function is, at least in a first step, obtained by solving a set of smaller single electron wave functions (or orbitals) for each of the electrons in a system. An example of an electron-molecule reaction is Dissociative Electron Attachment (DEA) which plays a major role in different technological, atmospheric and biological environments. In it, a molecule dissociates into fragments after capturing a free electron. A similar type of reaction is Dissociative Electron Transfer (DET) where instead of capturing a free electron, an electron is transferred from an atomic or molecular species to another, leading to fragmentation of the molecule receiving the electron. In chapter one we describe the theory behind several ab initio methods that are fundamental for studying important electron-molecule reactions such as DEA which plays an important role in various biological processes. In Chapter two we discuss molecular force fields and their importance in computational quantum chemistry, and we also describe three families of force fields commonly used in the literature. In chapter three a brief description of radiosensitizers and their application in both biology and medicine is given. In chapter four the essential theory behind the methods employed to calculate Electron-impact Ionization Cross-Sections (EICS) is outlined. EICSs that are needed in many fields, often connected to plasma science, such as medically used micro-plasmas, lighting devices and semiconductor manufacturing. It is also connected to the modeling of effects of impurities in nuclear fusion reactors. Those impurities are small molecules inside a hot plasma, formed by erosion of the walls. In chapters five through seven we present three works that have been published in peer-reviewed journals through the course of my doctorate studies. Finally, some conclusions based on the results obtained in the papers published for this dissertation are drawn, and some future perspectives are given.
- Advancements in Nd:YAG Solar Laser Stability Under Solar Tracking Error ConditionsPublication . Catela, Miguel Trindade; Liang, Dawei; Vistas, CláudiaSolar-pumped lasers represent a promising frontier in laser technology, offering the ability to convert incoherent solar radiation into coherent laser radiation without relying on traditional pump light sources. This dissertation explores methods and strategies to develop and adapt solar-pumped laser technology, tailoring it to meet both scientific and industrial challenges. Firstly, solar laser configurations are applied to lamp-pump lasers commonly utilized in indus- try, with the objective of boosting efficiency and brightness while addressing thermal concerns. Additionally, the investigation of alternative beam profiles, such as top hat and doughnut- shaped, showcases the potential of solar lasers in diverse applications, including material pro- cessing and biomedical imaging. Moreover, efforts to improve the stability and efficiency of solar-pumped laser systems, particularly under solar tracking error condition, are discussed. By utilizing end-side pump configurations and incorporating a solar flux homogenizer, significant advancements have been made towards achieving uniform and stable solar laser emissions. Experimental validation of these advancements at the PROMES-CNRS solar facility in Odeillo, France, demonstrates the feasibility of building stable solar laser systems with low-cost solar tracking systems. Even when the tracking system was turned off, the total output power ex- tracted from the solar-pumped laser remained stable for 1 min, representing, to the best of our knowledge, the first successful demonstration of stable multibeam solar laser operation without solar tracking.