Coronary Artery Disease is one of the leading causes of death worldwide. Computed Tomography and Coronary Computed Tomography Angiography are the gold standard techniques for Coronary Artery Disease diagnosis. Some recent studies have found a connection between Coronary Artery Disease occurrence and the accumulation of visceral adipose tissue in the ventral cavity. By performing ventral cavity segmentations in Computed Tomography scans, it is possible to quantify and analyze important textural characteristics of visceral fat. However the manual delineation of this structures is a time consuming process subject to variability. An automated process would achieve a faster and more precise solution. This paper explores the use of a U-Net architecture to perform ventral cavity segmentations. Experiments with different input image sizes and types of loss functions were employed. The model with the best performance achieved a 0.974 Dice Score Coefficient which is a competitive result when compared to the state of the art methods.

Radiomics refers to the extraction of hand-crafted features from radiographic images. Combined with machine learning and data analysis algorithms, it can provide a valuable tool to enable a phenotypic tumour profile. Based on the hypothesis that quantitative analysis of medical images may have a similar prognosis power to phenotypes and gene protein signatures, radiomics debates with the lack of standardisation and reproducibility issues. CT convolution kernels modify the frequency contents of projection data before back projection during image reconstruction, affecting the values of, mainly, intensity and texture features. This study evaluated the effect of eight convolution kernels from two General Electric (GE) Computed Tomography (CT) scanners on 19 patients. Feature extraction was restricted to the Clinical Target Volume (CTV), manually defined by experts at Instituto Português de Oncologia do Porto Francisco Gentil (IPO-PORTO). Afterwards, the feature set was grouped per patient following the variance computation kernel-wise. Results show that shape-based features are invariant to changes in the convolution kernel, while Gray Level Size Zone Matrix (GLSZM) and GrayLevel Run Length Matrix (GLRLM) seem more exposed to such changes. Additionally, results also suggest that first-order features can withstand slight modifications to the kernel, much like Gray Level Co-occurrence Matrix (GLCM).

The generation of synthetic data is often used as a data augmentation technique for training deep learning models. In this work, we investigate whether synthetic medical datasets obtained through generative adversarial networks contain identifiable characteristics of the training data, threatening patient privacy. We propose various methods to classify a set of images as having been used or not used in the training of the model that originated a set of synthetic images. The empirical results support the hypothesis that synthetic data compromises the privacy of patients in the training data and, thus, should be subjected to the same regulations as real data when used in real-world clinical applications.

With the ever-growing complexity of deep learning models for face recognition, it becomes hard to deploy these systems in real life. Researchers have two options: 1) use smaller models; 2) compress their current models. The usage of smaller models might lead to concerning biases. However, compressing might be also responsible for an increase in the bias of the final model. We investigate the racial bias of a State-of-the-Art quantization approach when used with synthetic and real data. This analysis provides a few more details on the benefits of performing quantization with synthetic data, for instance, the reduction of biases on test scenarios. We tested four distinct architectures and evaluated them on a test dataset which was collected to infer and compare the performance of face recognition models on different ethnicity.

The automatic extraction of concepts and clinical descriptions from medical images may facilitate the work of clinicians. Besides, it may also contribute towards an increase in the trust of clinicians in artificial intelligence methods since their learning depends on structured clinical information. In this work, we develop and compare various approaches to detect concepts and generate reports (i.e. perform image captioning) from medical images. Regarding concept detection, we explored multi-label classification, adversarial training, autoregressive modelling, image retrieval, and concept retrieval. We also developed three model ensembles merging the results of some of the proposed methods. For the caption prediction task, we developed language generation models and compared them with a simple approach based on image retrieval.

Music can convey fundamental emotions like happiness and sadness and more intricate feelings such as tenderness or grief. Understanding the neural mechanisms underlying music-induced emotions holds promise for innovative, personalised neurorehabilitation therapies using music. Our study investigates the link between perceived emotions in music and their corresponding neural responses, measured using fMRI. Fifteen participants underwent fMRI scans while listening to 96 musical excerpts categorised into quadrants based on Russell's valence-arousal model. Neural correlates of valence and arousal were identified in neocortical regions, especially within music-specific sub-regions of the auditory cortex. Through multivariate pattern analysis, distinct emotional quadrants were decoded with an average accuracy of 54% +-16%, surpassing the chance level of 25%. This capacity to discern music's emotional qualities has implications for psychological interventions and mood modulation, enhancing music-based treatments and neurofeedback learning.

We propose two bilinear models to detect Gastric Intestinal Metaplasia (GIM) in narrow-band images that combine the embeddings of a pre-trained Deep Neural Network (DNN) with the outcome of a local texture descriptor based on fractal geometry. Our methods improve the DNN performance by a significant margin over several metrics (e.g., area under the curve (AUC) 0.815 vs. 0.738) in a dataset comprised of EGD narrow-band images.

Dialogue State Tracking (DST) is a technique that monitors the current state of a conversation by keeping track of filled slots and understanding the user's most recent actions within the dialogue. In this paper, we explore various approaches to DST in task-oriented dialogues. Given that there were no existing task-oriented datasets in Portuguese, we created the first such dataset - MultiWOZpt - which was inspired by the widely - recognized MultiWOZ dataset. Among the approaches we tested, the best results were achieved using BERT-base and BERT-large models, and the choice between the two will depend on the specific task the user wishes to solve. These models demonstrated significant promise for task-oriented dialogue systems in Portuguese.

Breast cancer is a significant global health concern, affecting thousands of individuals, primarily women, with estimated cases expected to climb by 2040. This paper describes the creation of an application to collect ground truth data to aid engineers in the development of a mammography retrieval system. The application is built upon OutSystems, a low-code application platform. Key features of the application include allowing experts to view probe images and associate them with relevant images from the database. Additionally, the platform allows image filtering based on eigth mammogram dimensions. While the ultimate goal is to create a system for medical specialists, the current platform represents a step in the process, facilitating the acquisition of ground truth.

Pain is a highly subjective and complex phenomenon. Current methods

measure pain mostly rely on the patient’s description, which may not always be possible. Thus, pain recognition systems based on body language and physiological signals have emerged. As the emotional state of a person can also influence the way pain is perceived, in this work, a protocol for pain induction with previous emotional elicitation was conducted.

Eletrocardiogram (ECG), Electrodermal Activity (EDA), and Eletromyogram (EMG) signals were collected during the protocol. Besides the physiological responses, perception was also assessed through reported-scores (using a numeric scale) and times for pain tolerance and threshod. In this protocol, three different emotional elicitation sessions (negative, positive and neutral) were performed.

The results showed that during the negative emotional state, pain reported-scores were higher and pain threshold and tolerance times were smaller when compared with positive. As expected, the physiological response to pain remain similar despite the emotional elicitation.

Pain is highly subjective and difficult to quantify. The current methods used to assess pain request that the person is able to communicate which is not always possible. This way, automatic classification of pain has emerged, namely using physiological signals. Quantitative sensory testing (QST) is a safe way to induce pain. In this work, a protocol for inducing pain using a thermal QST device was implemented. Three levels of pain (low, medium and high) were calibrated and then induced while recording physiological signals. The results show that features obtained through Electrocardiogram (ECG) and Electrodermal Activity (EDA) could differentiate between pain levels and non-painful states, but could not differentiate well among pain levels.

Data storage in synthetic DNA is being considered as a solution to the problem of preserving the ever increasing digital information being created by humans. Unfortunately the biochemical processes involved in this type of storage result in chemical degradations and errors that need to be characterized to devise measures able to minimize the loss of information. Since the cost of synthesising, storing and sequencing DNA are high, research in DNA storage has to use simulators that can emulate those processes, simulating the errors and degradations that occur in real-life situations.

In this paper we describe some experiments with two DNA storage simulators, showing how they are used, indicating their best use scenarios, and reporting their performance in terms of simulation accuracy.

The detection of network flows carrying malicious data has been left to a set of techniques and tools based on machine learning (ML), with the disadvantage of requiring vast computational resources and sometimes low generalisation capacity when faced with new types of attack, known as a zero-day attack.

This paper describes the application of a model based on three statistical natural laws, Benford’s, Stigler’s and Zipf’s Laws, as a model for detecting malicious flows extracted from network traffic, as well as the results obtained from a dataset with 40000 network flows, where 20000 are classified as malicious flows and the remaining as benign flows. To classify network flows as malicious or benign, three statistical tests of different natures were used: parametric (Pearson and Komolgorov correlation and p-value calculation) and non-parametric (Cramer-Von Mises p-value calculation), applied to the results of the frequency of occurrence of the first digit with the empirical frequency of each natural law.

Although the results obtained with the model based on the laws of Benford, Stigler and Zipf do not surpass the results obtained by the majority of models based on machine learning, as initial results, we emphasise that they are satisfactory, with a maximum F1 of 69.40% having been obtained.

Regularization techniques enhance deep learning models' generalization and robustness. We evaluate Patch Mix for Transformers. Patch Mix involves random patch replacing within the input sequence with similar patches of a Transformer during training. This encourages the model to learn more robust representations independent of patch locations, avoiding co-adaptation, improving generalization, and mitigating overfitting. Evaluating Patch Mix on benchmark datasets, we compare it with Dropout. Results show Patch Mix effectively reduces overfitting, focusing on meaningful patch interactions rather than specific locations.

Modern reinforcement learning is a field its with many dynamic and fruitful integration with various engineering and scientific disciplines. However, it comes with an inherent challenge—the understanding of its models, which makes it challenging for humans to trust the decisions made by these algorithms. In response to this challenge, we propose the development of an interactive dashboard designed to ease the analysis of reinforcement learning environments. This dashboard offers a comprehensive set of features to visualize critical elements of reinforcement learning experiments, encompassing agent behavior, reward dynamics, and exploration of time-based features.

Deep learning methods show a high performance in different problems. However, when the target domain has a distribution shift from the source domain, a significant reduction in the performance of the models occurs. Learning models applied to medical imaging analysis are limited by characteristics of the medical datasets, which are typically small, low representative and not completely annotated, resulting in problems of domain shift. Typically, these problems are addressed by strategies for domain generalization or adaptation. The literature shows the need to create learning models with stronger generalization capability to deal with two problems: lack of generalization (P1) for cohort populations and (P2) for distinct imaging modalities. This paper gives a perspective on how to address these two problems, mainly centred on the importance of strategies to force the learning of domain invariant feature representations.

The field of industrial automation is undergoing a rapid transformation, driven by the integration of artificial intelligence (AI). Among the most innovative applications of AI in this field is its remarkable ability to convert hand-drawn images of application drafts into functional code.

Drafts can be from new or old projects, paper or digitally acquired. This groundbreaking advancement serves as a bridge between design and execution, enabling more flexible and efficient programming. In this paper, we will delve into the potential and impact of this AI-driven process on industrial automation.

In this paper, a study on the stability of three deep learning point cloud compression solutions, notably ADLPCC, PCC GEO CNNv2, and PCGCv2 is presented. This study aims to show one of the problems of deep learning based codecs, that results in instability on the codec performance. These behaviours appear in unusual cases. The same architecture with a new training will most likely solve the problem. An example of a point cloud representing cultural heritage is used in this paper for demonstration purposes.

Three different training sessions were conducted using the default training set and cost function of each of the considered codecs. Across each epoch of the training sessions, the objective quality metric MSE PSNR D1 was computed. The final result of each training session was objectively to the default implementation of the codecs, using the MSE PSNR D1 and PCQM metrics.

Periodic inspection of electrical grid assets is fundamental to efficiently manage maintenance activities and guarantee the safe and reliable operation of the grid. For efficient visual inspection, given the huge amounts of collected image data, automated inspection is mandatory. In this paper we investigate a deep learning approach to the automated classification of defects in operating photovoltaic modules through the analysis of outdoor collected thermography images. We test several variants of a simple deep learning architecture on a publicly available dataset of labeled infrared images divided into 12 classes. The classification results yield Precision, Recall and F1-score of 0.87, 0.86 and 0.86, respectively, demonstrating the practical value of the presented automated approach.

Electric grid asset inspections are critical to ensure that modern society can continue to depend on a stable and uninterrupted supply of electricity. This paper represents an initial investigation to test the efficiency of a deep learning model to address the problem of identifying defects on visible light images of power line insulators. We tested a publicly available dataset with 1688 images containing over 6000 shell insulators, using a Faster R-CNN architecture on the Detectron2 framework. We achieved a mean average recall (mAR@50:95) of 82,6% with a 79,9% mean average precision (mAP@50:95) for the three considered classes.

Effective pain management depends on an accurate assessment of pain intensity. However, limitations in current pain assessment scales, including subjective reporting of pain and potential observational bias, can result in inadequate pain treatment. Therefore, there is increasing interest in the development of objective pain assessment methods, particularly employing physiological indicators, with the goal of improving pain assessment. The aim of this work was to classify pain levels, induced by a Cold Pressor Task (CPT), using features extracted from electrocardiogram (ECG) signals. The Random Forest algorithm demonstrated superior performance in distinguishing between higher (NRS>8) and lower (NRS≤8) levels of pain, using a set of 15 ECG-features associated with the P, R, S, and T waves. This model achieved an accuracy of 95.3%, an F1-score of 94.0%, a precision of 97.9%, and a recall of 90.4%. These results demonstrate the feasibility of using physiological variations in the ECG signal for assessing pain.

The accurate reconstruction of viral genomes has become increasingly important due to the growing availability of diverse human viral sequenced samples. This necessity is particularly pronounced in clinical and forensic scenarios, where specialized tools capable of correctly reconstructing genomes are required. In this article, we present CoopPipe, a method capable of improving the reconstruction of human DNA viruses. CoopPipe utilizes and adaptive cooperation between publicly available viral reconstruction tools and is capable of improving the reconstruction process by, on average, 19.3% in terms of the NCSD and 17.0% in terms of the NRC, in relation to the best tool and for each virus. The implementation of CoopPipe is entirely reproducible and publicly available at https://github.com/viromelab/CoopPipe.

Optimizing radiation doses for computed tomography (CT) scans is essential to ensure patient safety by minimizing potential health risks associated with ionizing radiation exposure. The necessity arises from the key role of CT scans in disease diagnosis and monitoring, balanced against the inherent radiation hazards they pose. To address this need, an innovative project is underway, proposing the application of Artificial Intelligence (AI) models to predict ionizing radiation doses within CT scans, effectively refining radiological practices and minimizing associated risks. The project aims to enhance healthcare quality by aligning doses with Diagnostic Reference Levels (DRLs) in Portugal. The approach employed is personalized, adapting imaging protocols based on individual patient attributes, following the ALARA (As Low as Reasonably Achievable) principle, aimed at achieving the minimal radiation exposure necessary for diagnosis without compromising accuracy. By leveraging technology, the project empowers healthcare professionals with efficient data analysis, facilitating precise diagnoses and effective treatments. This initiative contributes to compliance with radiological protection guidelines, contributing to Portugal's commitment to radiological protection, while healthcare professionals diligently adhere to medical protocols. These collaborative actions prioritize patient welfare and improve healthcare safety by mitigating the risks of ionizing radiation and promoting a technologically advanced healthcare environment.