APRÈN. Portal de la producció docent del professorat de la UPC

Virtual reality is an emerging technology that is having a strong impact across various fields, especially in the world of video games. Although the theme of climbing has been explored in some commercial titles, there is currently no proposal that faithfully combines the immersive experience of virtual reality with a realistic representation of the climbing environment. The objective of this project is to develop a virtual reality climbing video game that reproduces, as accurately as possible, the conditions faced by a real climber. To achieve this, factors such as fatigue, difficulty depending on wall inclination, and various technical elements that affect the practice of this sport have been taken into account.

Aquest treball se centra en la verificació automatitzada del compliment de normatives de construcció en edificis mitjançant l'ús de grafs i models d'intel·ligència artificial. L'objectiu principal és desenvolupar un sistema capaç d'identificar possibles incompliments de normatives en edificis, basant-se en la transformació d'arxius IFC i normatives legals en grafs estructurats. El sistema proposat combina models de processament del llenguatge natural, visió per ordinador i bases de dades de grafs per analitzar edificis i normatives, integrant aquestes dades i contenidors en un entorn global i eficient. No obstant això, no només es crea un producte mínim viable funcional i eficaç, sinó que també s'estableixen les bases fonamentals per a la construcció d'un sistema més complet i avançat en el futur. S'ha realitzat un estudi profund dels components necessaris, com interactuen entre ells i com poden evolucionar per garantir que aquest prototip inicial pugui ser escalable i robust en aplicacions reals més àmplies. Aquesta estructura assegura que el treball no es limiti a un simple prototip, sinó que serveixi com a base sòlida per a futurs desenvolupaments. Tot i les limitacions de recursos disponibles durant el desenvolupament, el sistema ha demostrat ser efectiu en les seves tasques. Tanmateix, amb més poder computacional, es podria millorar significativament la qualitat dels resultats, aconseguint un grau més alt de precisió i eficiència. Aquest MVP, juntament amb les bases assentades, marca un primer pas cap a un sistema capaç de transformar la verificació de normatives en la construcció, millorant la seguretat, la sostenibilitat i la simplicitat dels processos.

This thesis addresses uncertainty in automated brain tumor segmentation using deep learning. An ensemble of SwinUNETR, SegResNet, and Attention U-Net models was developed, incorporating test-time augmentation and test-time dropout for uncertainty estimation. Evaluation on the BraTS 2021 dataset showed that the ensemble achieved statistically significant improvements in segmentation accuracy, particularly for edema, relative to individual models. Uncertainty maps correlated with voxel-wise error and enabled risk-aware prediction. A Streamlit application was prototyped for clinical feasibility, allowing for interactive visualization of segmentation and uncertainty. The developed application lays the theoretical and technical groundwork for future compliance studies in medical image segmentation, with the potential to accelerate regulatory approval and clinical integration. Further progress will require larger, more diverse models and prospective reader studies to translate these preliminary findings into clinically meaningful improvements.

Advances in Assisted Reproductive Technology (ART) have underscored the need for objective, data-driven tools to support embryo selection in In Vitro Fertilization (IVF). This thesis explores image-based artificial intelligence methods for predicting good-quality blastocyst formation at day 3 using time-lapse data from EmbryoScope incubators. A custom dataset of 10,274 annotated embryo sequences, collected in collaboration with Hospital Clínic de Barcelona, was assembled and preprocessed for model training. Multiple approaches, including traditional machine learning and deep learning models, were evaluated across different modalities: single images, time-lapse sequences, and fused morphokinetic features. The best-performing model - based on a EfficientNetB0 and attention-based feature fusion - achieved a balanced accuracy of 77.37% on the test set, while a fully automated video-based model incorporating an LSTM network reached a balanced accuracy of 75.3%. The results demonstrate a strong performance of the embryo models, reaching state-of-the-art results on the dataset at hand and showing promising potential for real-world clinical support in embryo selection.

Este Trabajo de Fin de Grado, desarrollado en el marco del grado en Ingeniería Informática en la Facultad de Informática de Barcelona, se centra en la creación de una herramienta de segmentación de pinturas murales basada en técnicas de inteligencia artificial, en concreto en la IA de Meta SAM (Segment Anything Model), una IA centrada en la segmentación de objetos en imágenes. Este proyecto afronta el reto de ofrecer una herramienta web sencilla y versátil con la capacidad de segmentar regiones de pinturas murales basándose en figuras, zonas de deterioro, pigmento u otras características mediante el uso de una combinación de herramientas simples. Este proyecto está dirigido a historiadores del arte que puedan necesitar esta herramienta de segmentación para el estudio, el análisis o la restauración de pinturas murales.

Despite the increasing integration of deep learning in medical imaging, the limited availability of annotated datasets remains a critical bottleneck, particularly for emerging conditions such as Long COVID. Traditional data augmentation techniques often fail to capture the complexityof3DbrainMRIstructures, leading to suboptimal model generalization. In this study, we propose a slice-based latent diffusion framework designed to synthesize 3D brain MRI volumes in a computationally efficient slice-by-slice manner. By modeling the joint distribution of MRI slices and incorporating positional embeddings, our approach maintains spatial coherence across volumes while significantly reducing the computational burden typically associated with 3D diffusion models. Furthermore, the model is conditioned on diagnostic labels, enabling the generation of class-specific synthetic data for both healthy individuals and Long COVID patients. Experiments on classification tasks demonstrate the effectiveness of the generated synthetic data for augmenting small, imbalanced datasets, ultimately improving the diagnostic performance of deep learning models in the context of Long COVID classification.

The Facultat d'Informàtica de Barcelona (FIB) is an institution known for its knowledge in computer science, using the knowledge learned in this degree, this project aims to develop an artificial intelligence capable of competing in turn-based games, such as Pokémon, using reinforcement learning techniques. The specialization of computing helps the students to possess the knowledge and skills to design adequate algorithmic strategies, such as Monte Carlo or Minmax which will be very helpful in developing bots that will train the AI. To fully understand the scope of this study, it is essential to define the terms and concepts related to artificial intelligence, machine learning and reinforcement learning. We define AI as the study of agents that receive percepts from the environment and perform actions. Each such agent implements a function that maps percept sequences to actions, and we cover different ways to represent these functions, such as reactive agents, real-time planners, and decision-theoretic systems. [1] Machine learning is essentially a form of applied statistics with increased emphasis on the use of computers to statistically estimate complicated functions and a decreased emphasis on proving confidence intervals around these functions. [2] Reinforcement learning is learning what to do—how to map situations to actions—so as to maximize a numerical reward signal. The learner is not told which actions to take, but instead must discover which actions yield the most reward by trying them. In the most interesting and challenging cases, actions may affect not only the immediate reward but also the next situation and, through that, all subsequent rewards. These two characteristics—trial-and-error search and delayed reward—are the two most important distinguishing features of reinforcement learning.

Glioblastoma is currently considered one of the most aggressive brain tumours, characterised by rapid progression, high heterogeneity, limited treatment success, and frequent recurrence. In this work, we propose a machine learning pipeline to predict treatment outcomes in a preclinical mouse model of glioblastoma using longitudinal MRI data. Our approach first applies classical machine learning techniques to radiomic features extracted from segmented tumour regions, with the aim of identifying imaging biomarkers that correlate with biologically meaningful patterns of tumour progression or regression. We then introduce a complementary deep learning strategy based on convolutional neural networks, which learns hierarchical representations directly from the images and reduces reliance on manual feature engineering. Both pipelines are rigorously evaluated and compared. The overarching goal is to discover early predictors of tumour relapse or cure that align with underlying biological processes, contributing to the development of non-invasive monitoring strategies and enhancing treatment assessment in translational research.

En una era profundamente conectada, donde las comunidades digitales se vuelven cada día más predominantes, es crucial entender cómo los procesos algorítmicos de recomendación y agregación de contenidos en plataformas como X (antes Twitter) pueden favorecer el desarrollo de cámaras de eco, la segmentación ideológica y, en última instancia, la radicalización de los discursos. Desde las implicaciones sociales hasta el análisis computacional, este estudio explora la topología de las comunidades de internet y el fenómeno de las redes sociales de polarización ideológica a través de un enfoque interdisciplinario que reúne técnicas de análisis de redes y comprensión sociológica crítica. Técnicamente, el proyecto implementa y compara tres de los algoritmos más populares de detección de comunidades: Louvain, Leiden y Scalable Community Detection (SCD). Se aplican a diversas redes sociales, con el fin de examinar cómo varían las particiones obtenidas en función del algoritmo empleado. La investigación encuentra diferencias aparentes en el tamaño, el número y la cohesión de las comunidades detectadas, con implicaciones para la interpretación de la polarización dentro de las redes. En cuanto a la medición de la polarización, se ha implementado y utilizado la métrica propuesta por P.H.C. Guerra basada en la forma de los límites entre comunidades. Ofrece una mejor opción que la modularidad para medir el nivel de interacción (o falta de interacción) entre los grupos de la sociedad. Sin embargo, se han descubierto limitaciones aún más severas con respecto a la escala y la aplicabilidad a redes que involucran a numerosas comunidades. Para ello, se propone un enfoque mejorado para el cálculo de la polarización global de la red a través de un proceso de fusión iterativa de comunidades y un cálculo optimizado. Además del análisis computacional, el artículo presenta un punto de vista sociológico sobre cómo las redes sociales funcionan como plataformas de expresión y cristalización de movimientos populistas. Se presta atención a la interconexión de la polarización, el populismo y la radicalización, y a cómo estos procesos se ven intensificados por la naturaleza algorítmica de las propias plataformas en línea. La investigación muestra que todos estos fenómenos deben abordarse con una mentalidad crítica y multidisciplinaria, combinando ingeniería y sociología para desarrollar herramientas más justas y conscientes para examinar y luchar contra la fragmentación social en línea.

Purpose: The primary purpose of this study is to develop deep learning (DL) frameworks capable of assessing the progression of diabetic retinopathy (DR) in patients with Type 1 diabetes mellitus (DM1) over a 5-year span (±2 years), addressing a significant gap in the current literature and enhancing the potential for early detection in ophthalmological clinical applications. Data: The dataset utilized in this research comprises Optical Coherence Tomography Angiography (OCTA) images collected from a clinical trial conducted at the Hospital Clínic of Barcelona. This dataset includes images from 171 participants, focusing on patients with DM1, featuring varying degrees of DR progression over a 5-year period (±2 years), along with multiple image resolutions (3×3 mm and 6×6 mm) and capillary plexus depths - superficial capillary plexus (SCP) and deep capillary plexus (DCP). Methods: The data pipeline begins with image concatenation to create a unified fourdimensional (4D) OCTA input, combining 3×3 mm and 6×6 mm scans across two depth levels - the SCP and DCP - followed by min-max normalization. To address class imbalance, a combination of techniques was employed, including undersampling, class weights, batch balance, and data augmentation, both individually and in combination. Transfer Learning (TL) was applied using DenseNet201, Inception-V3, and VGG19, supervised learning (SL) models, chosen to leverage existing knowledge from ImageNet dataset. The main metrics used are F1-Score (F1), Accuracy (ACC) and Area Under the Curve (AUC). Results: The best-performing models, obtained through 5-fold cross-validation, are DenseNet201 (F1: 64.50%; ACC: 77.50%; AUC: 82.03%), with ACC per class of 53.33% for Better, 60.00% for Worse, and 83.72% for No Change, and Inception-V3 (F1: 58.03%; ACC: 71.79%; AUC: 76.55%), with per-class ACCs of 46.67% for Better, 48.57% for Worse, and 71.78% for No Change. DenseNet201 employed entry layers for detail extraction and added top layers above the base model, using a balanced batch size. In contrast, Inception-V3 added only top layers and applied data augmentation via cropping to 100% of the data, while also incorporating class weights to address imbalance. Conclusion: This study was the first to focus exclusively on the assessment of the progression of DR on DM1 patients, addressing a critical gap in the literature. Despite limitations in dataset size and resolution, it demonstrated that DenseNet201 and Inception-V3 are effective models when combined with TL and robust imbalancehandling techniques such as class weights, batch balance, and data augmentation - successfully minimizing misclassifications, particularly among patients likely to experience worsening.

Anatomical magnetic resonance imaging (MRI), including T1 and T2 weighted images, is routinely used to visualize anatomical structures and detect pathological changes, as these imaging techniques provide complementary contrasts that highlight distinct tissue properties and abnormalities. Diffusion tensor imaging (DTI), a more specialized MRI technique, enables the mapping of white matter and the study of structural connectivity through metrics like fractional anisotropy (FA) and mean diffusivity (MD). However, DTI acquisition is time intensive and often excluded from standard clinical protocols. This study introduces a method to synthesize FA and MD images from widely available T1 and T2 weighted anatomical scans, reducing dependence on resource intensive DTI acquisition. Radiomics, a rapidly evolving field, focuses on extracting quantitative features from medical imaging, to identify biomarkers associated with clinical labels. It is a well established tool in oncology, particularly for tumor segmentation. Fully convolutional neural networks (FCNNs) are also employed for similar tasks, such as deriving clinical labels for tumor segmentation. However, their reliance on computationally intensive 3D convolutions often necessitates the use of 2D slices from 3D volumes, limiting their feasibility. We present a hybrid approach that combines the strengths of radiomics and neural networks. Specifically, we use a feedforward neural network (FNN) as a classification or regression head applied to radiomic features, preserving 3D spatial information during feature extraction while reducing the computational demands of 3D convolutions. This neural network head, akin to fully connected layers in traditional convolutional neural networks (CNNs), offers significant flexibility. Our experiments are limited to the basal ganglia, a crucial region involved in brain structure and function that is significantly impacted by neurodegeneration in Huntington's disease (HD). The inclusion of HD patients allows for assessing the method's reliability and robustness. Additionally, we explore the potential of using the T1/T2 ratio as an input image, which has been proposed in recent studies, as a proxy for the myelin content of the brain. Since myelin plays a role in how DTI functions, the T1/T2 ratio may enhance model performance, effectively bridging anatomical MRI and DTI. The reported results of the experiments indicate strong performance, with Pearson correlations of 0.85 for FA and 0.95 for MD predictions. While the correlation decreases significantly for FA and moderately for MD in HD patients, the model's performance remains consistent when both healthy controls and patients are analyzed together. These findings underscore the promise of our hybrid approach for synthesizing structural connectivity images and improving accessibility to diffusion metrics.

Autonomous driving systems are poised to revolutionize transportation, enhancing safety, efficiency, and accessibility. However, adverse weather conditions, such as heavy rain, dense fog, snow, and nighttime darkness, remain critical challenges, often compromising the performance of object detection systems supporting the perception of vehicles. This thesis aims to address these challenges by developing novel techniques for enhancing the robustness of object detection systems under adverse weather conditions, using state-of-the-art deep learning models and innovative data augmentation and denoising strategies. The research leverages diffusion models, specifically the Instruct Pix2Pix variant of Stable Diffusion, to simulate realistic adverse weather conditions on existing images and with that perform a data augmentation and to develop a specialized denoising model to mitigate weather-induced distortions. Using both simulated data from the CARLA simulator and real-world datasets such as BDD100K and ACDC, this study evaluates the performance of widely used object detection models, including Faster R-CNN and YOLOv10, with the aim of first analyzing and investigating whether the developed techniques are effective in the simulated environment and subsequently validate them in real-world scenarios. Key contributions include the creation of augmented datasets incorporating synthetic adverse weather conditions and the development of a denoising model to improve image clarity. The results show that the proposed data augmentation strategies enhance model robustness under extreme weather conditions, paving the way to further explore this technique to make models robust against weather-related factors and other challenges such as variations in object colors or pedestrian attributes. However, despite the promising results in the simulation environment, the denoising model has proven to be ineffective in real-world scenarios: it fails to remove the adverse weather effects and instead deteriorates the image quality, highlighting the complexity of real-world data and the need for more sophisticated denoising approaches. This study advances the field of autonomous driving by providing solutions to improve object detection performance under adverse weather. The findings of this thesis contribute to the development of safer and more reliable autonomous vehicles by addressing a critical obstacle to real-world deployment of autonomous driving technology. Furthermore, this work offers valuable insights for future research in autonomous mobility and computer vision.

Aquest projecte presenta CatMol, un model generatiu dissenyat per crear petites molècules representades com a grafos, utilitzant matrius de nodes i arestes amb atributs categòrics. S'aplica un model probabilístic de difusió discreta a aquestes representacions, aprofitant una distribució marginal derivada del conjunt de dades, la qual millora el rendiment en comparació amb una distribució uniforme. Per assegurar que el model aprengui correctament la distribució del grafo, s'empra un mecanisme d'atenció que incorpora la informació de les arestes. El model s'entrena amb diversos conjunts de dades i es valora utilitzant diferents mètriques de rendiment. En conjunts de dades de molècules similars a fàrmacs, CatMol demostra un rendiment comparable amb els models més avançats, amb resultats superiors en alguns casos. A més, el rendiment del model millora amb conjunts de dades més grans, cosa que suggereix la seva escalabilitat. El treball futur inclourà entrenar amb conjunts de dades encara més grans per validar més la seva escalabilitat i el seu potencial.

Language models for RNA and proteins have revolutionized biological research, yet the principles governing their scaling remain an underexplored frontier. While numerous studies continue to investigate new architectures and training strategies, a systematic understanding of how scaling laws influence their performance is still lacking. In this work, inspired by the scaling methodologies proposed by Hoffmann et al. [1] for Large Language Models we systematically explore different configurations of protein and RNA language models, with a particular emphasis on tokenization strategies and the interplay between model parameters and training tokens. Our findings reveal that, while both parameter count and dataset size contribute to model performance, the number of training tokens exerts a more profound influence. In the case of pLMs, we successfully broke the performance plateau observed in ESM by developing MMFreePLM, an architecture that achieves superior efficiency while reducing computational overhead. This demonstrates that architectural optimization can unlock substantial gains in scalability without necessitating an exponential increase in resources. Similarly, for rnaLMs, we show that ChaRNABERT, with only 50 million parameters, achieves performance on par with RiNALMo, a significantly larger model with 650 million parameters. These findings highlight a key insight that optimizing architectural design and tokenization strategies has a greater impact on biomolecular language model performance than merely increasing model size. By bridging critical gaps in the understanding of scaling laws for pLMs and rnaLMs, this study paves the way for the development of more computationally efficient and scalable biomolecular models. Our work not only enhances the feasibility of large-scale biological sequence modeling but also contributes to advancing biotechnological research aimed at decoding the language of life with greater precision and efficiency.

This thesis investigates the optimization of user communication patterns on Leboncoin, France's leading second-hand marketplace platform, through the implementation of context-sensitive message suggestions. The research addresses a critical business challenge: approximately 23% of conversations on the platform receive no initial response, a phenomenon known as “ghosting” that significantly impacts user experience and platform revenue. While previous experiments on the Italian sister company Subito, demonstrated that fixed message suggestions could improve reply rates, this research explores showing context-sensitive full text suggestions. At the core, these suggestions are based on classifying the latest message received and then providing a set of potential, manually created replies depending on the class. Since a wide range of product categories are sold on our market places, we have a particular interest in automating the discovery of new and existing classes through unsupervised clustering. This master's thesis describes our attempts to cluster the messages based on different types of embeddings and compare the results obtained to the classic text representations TF-IDF (term frequency - inverse document frequency) and BOW (bag of word). Additionally, we attempt to further improve the clusterings through outlier detection, normalizing and applying PCA (principle component analysis) to the embeddings and combining multiple “views” of text representations in order to improve results with mixed success.

Nowadays data are in motion, change continuously and are –potentially– unbounded implying data sources that are also in constant evolution. From the point of view of data persistence, this reality breaks the usual paradigm of dynamic although stable data sources. Besides, the number of applications to help critical decision making in real time is also rapidly increasing. These two scenarios raise the need of re-thinking both the data and the query models to fit these new requirements. So that, under these circumstances, it seems that a continuously evolving data graph is a suitable data model to use and therefore to study and analyze. Thus, in this work, we tackled the problem of querying continuously evolving data graphs in a specific context: the context of ATM1 transactions, in particular anomalous ones. Under this context, evaluating continuous queries corresponds to recognizing patterns –usually associated with anomalous behaviors– in the volatile (evolving) subgraph of ATM transactions. To do so, we propose an evaluation process based on the so called dynamic pipeline computational model, a stream processing technique that facilitates the emission of alerts as soon as anomalous patterns are identified. Stream based bank applications that monitor ATM transactions are direct beneficiaries of our proposal since they can continuously query data graphs to get “fresh” data as they are produced, avoiding the computational overhead of having to discard non-valid data, as current systems work.

Echo state networks (ESNs) are a type of recurrent neural network (RNN) consisting of a large, randomly connected recurrent reservoir followed by a linear readout layer. Training ESNs boils down to a linear regression because only the readout weights are adapted. Despite their conceptual simplicity, successfully applying ESNs can be challenging, as their behavior is not yet well understood from a theoretical point of view. This thesis intends to contribute to the understanding of these networks by exploring various ESN-related aspects. Our main focus lies on the role of the spectral radius, which is one of the most important hyperparameters of an ESN. We theoretically derive the relation between reservoir weight distribution and spectral radius and demonstrate the influence of the latter on the performance of an ESN for inputs of varying difficulty. Subsequently, we examine the effects of plasticity-based reservoir pretraining on the network's results and on its spectral radius and complete our experimental analysis with two prediction tasks on real-world data. Finally, we provide an outlook on an interesting alternative perspective on ESNs that relates these networks to kernel machines.

The integration of digital twins into urban planning, smart city development and infrastructure management has underscored their potential across a wide range of applications. However, significant challenges are still present in the creation of highfidelity digital representations, as well as their synchronization with the physical world and their efficient rendering on commonly used devices such as smartphones. This work proposes an architecture for the creation, adaptation and visualization of large-scale digital twins of urban environments, integrating data obtained from multiple heterogeneous devices. Furthermore, a mobile augmented reality application is being developed which demonstrates a use case based on the proposed architecture, showing promising rendering performance and visualization quality. In addition, various planned features are discussed, which can enhance the responsiveness and immersiveness of the digital visualization. Thus far, the application has been used to render digital twins of up to 180 million points in real time.

Brugada syndrome (BrS) is a rare, inherited cardiac condition associated with an increased risk of sudden cardiac death. Current risk stratification relies heavily on the presence of symptoms and invasive electrophysiological studies, particularly in asymptomatic individuals with characteristic electrocardiogram findings. This presents a significant clinical challenge in identifying high-risk individuals. This thesis explores the potential of Natural Language Processing (NLP) and Machine Learning/Deep Learning (ML/DL) techniques to improve BrS risk stratification using information extracted from Electronic Health Records (EHRs). The study develops and evaluates an NLP-driven pipeline to extract relevant clinical features from unstructured EHR data, including patient demographics, medical history, family history, medication use, and lifestyle factors. These features are then used to train and validate ML/DL models to predict the risk of BrS-related adverse events. The performance of various ML/DL algorithms is compared, and the most effective model is analyzed for its clinical interpretability and potential for real-world application. This research aims to provide a novel, non-invasive approach to BrS risk stratification, ultimately contributing to improved patient care and outcomes.

Deep learning models for weather and climate applications often demand massive datasets and extensive computational resources, creating bottlenecks in both training and inference. In this study, we propose a wavelet-based data compression strategy that, combined with a masked autoencoder (MAE), significantly reduces storage requirements, and increases GPU/TPU usage without sacrificing reconstruction quality. We decompose global ERA5 reanalysis images using a two-level Discrete Wavelet Transform, creating tiled wavelet subbands that can be selectively masked. The MAE is trained to inpaint missing detail coefficients while keeping the low-frequency approximation visible, efficiently leveraging spatial context from unmasked regions. Through experiments on resolutions of 5.625º and 2.8125º , we demonstrate that the finer-resolution models yield sharper reconstructions, and that a single-stage MAE outperforms a cascaded two-model pipeline for two-level DWT data. Furthermore, integrating these MAE reconstructions into a ResNet-based foundation model for forecasting shows that the proposed wavelet compression preserves essential meteorological features, maintaining competent predictive accuracy even when large portions of detail coefficients are masked. This ability to trade off between compression ratio and forecast skill, coupled with minimal overhead for data loading, underscores the potential of wavelet-based masked autoencoders to accelerate deep learning workflows in the atmospheric sciences and beyond.