This line applies Computer Vision, Machine Learning and Generative AI to dentistry. It develops intelligent systems for the automatic analysis of clinical images of the oral cavity, aiming to support early detection and computer-aided diagnosis of potentially malignant lesions and oral cancer.

This research is carried out in collaboration with the ORALMED research group at the Complutense University of Madrid (UCM), through OralMedNet, a telediagnosis service for oral mucosal lesions. Through a web application, dentists and physicians submit clinical images and lesion characteristics, which are reviewed by ORALMED's oral medicine specialists to improve the diagnosis of oral cancer and oral potentially malignant disorders. The project brings together dentists and computer scientists from UCM, UNED, UOC and other institutions.

Possible TFG/TFM projects

• Automatic classification of oral lesions with Deep Learning. Design and evaluation of neural networks that distinguish benign from potentially malignant lesions in clinical photographs.
• Automatic segmentation of lesions in intraoral images. Segmentation algorithms based on classical computer vision techniques or U-Net models.
• Explainability analysis (Explainable AI). Study of activation maps (Grad-CAM, Score-CAM, etc.) to interpret the decisions of medical classification models.
• Web application for diagnostic support. A prototype to view images, run inference and display classification results.
• Generative AI for synthesising oral lesion images. GANs and diffusion models that generate synthetic images to improve the training and validation of AI-assisted diagnosis systems.

Contact: Margarita Bachiller (marga@dia.uned.es)

This line explores how Artificial Intelligence and bio-inspired Computer Vision models can contribute to the development of visual neuroprostheses that improve the perception of the environment for people with visual impairments. This research is currently carried out within the AI4BRAIN project, in collaboration with the Universidad Politécnica de Cartagena (UPCT) and the Universidad Miguel Hernández (UMH).

UNED's contribution focuses on bio-inspired Artificial Intelligence techniques for image processing and retinal modelling, with the goal of generating more efficient visual representations for prosthetic vision systems (AI4BRAIN-UNED project).

Possible TFG/TFM projects

• Simulation of a cortical visual prosthesis based on bio-inspired encoding and stimulation–phosphene mapping. A simulator of a cortical visual prosthesis based on the Utah Array, studying the relationship between electrical stimulation patterns and phosphene generation to optimise visual perception.
• Optimisation of stimulation patterns for the Utah Array. Electrode activation strategies that improve phosphene generation and the quality of the visual perception provided by cortical neuroprostheses.
• Bio-inspired modelling of retinal processing. Simple models inspired by how the retina works, used for image preprocessing.
• Comparison of computer vision models for visual neuroprostheses. Evaluation of detection, segmentation and scene description models to find which provide the most useful information for prosthetic vision.
• Comparison of computer vision models for visual assistance. Performance evaluation of object recognition and semantic segmentation models in real-world scenarios.
• Automatic detection of relevant objects in a scene. Object detection models that identify elements of interest in everyday environments.
• Automatic scene description with Artificial Intelligence. Evaluation of vision-language models that generate scene descriptions to support the development of intelligent visual aids.
• Intelligent scene description for blind people. AI systems that identify the most relevant elements of a scene and produce representations adapted to users with visual impairments.

Contact: Mariano Rincón (mrincon@dia.uned.es)

This line investigates the development and application of Artificial Intelligence and Data Science techniques to support the early detection, diagnosis and monitoring of cognitive impairment and other neuropsychological disorders. Projects include the analysis of clinical data, traditional neuropsychological tests and immersive assessments in Virtual Reality, with the aim of building clinical decision support tools.

This line is currently carried out mainly through the SCREENIA project, a continuation of the IDENTIA project. Both focus on intelligent systems for the early detection of cognitive impairment through the analysis of digital biomarkers and Artificial Intelligence techniques. We also work on other projects that apply AI to the analysis of graphical, verbal, discourse and gestural neuropsychological tests, with the aim of obtaining assessment tools with greater discriminative and predictive power, able to detect the very earliest stages of cognitive decline while requiring fewer resources.

Possible TFG/TFM projects

• Analysis of tests performed in Virtual Reality. Study of metrics obtained during immersive cognitive tests, such as product selection patterns in a virtual supermarket, to characterise episodic memory strategies and develop digital biomarkers for the early detection of cognitive impairment.
• Automatic and explainable analysis of graphical tests. Computer vision and explainable AI methods to score drawing tests (for example, clock drawing or complex figures) against expert scores, and to extract digital biomarkers from the drawing process (stroke dynamics, timing, pauses).
• Automatic analysis of verbal fluency and spontaneous speech. Natural language and speech processing techniques to discriminate between normal ageing, mild cognitive impairment and dementia.
• Detection of patterns preceding clinical signs. Identification of early markers in people with subjective cognitive decline or risk factors, or through longitudinal follow-up, before recognisable clinical signs appear.
• Selection of the most informative test items. Feature selection techniques to identify the items or variables with the highest diagnostic value and design shorter or optimised versions of neuropsychological tests.
• Multimodal models for cognitive assessment. Models that combine graphical, verbal and gestural signals to improve the detection of cognitive impairment (recommended for TFM).

Contact: Estela Díaz (ediazlopez@dia.uned.es); Rafael Martínez (rmtomas@dia.uned.es)

This line applies Artificial Intelligence and Machine Learning to metabolomics in order to discover biomarkers that support the early diagnosis, prognosis and monitoring of disease. Metabolomics studies the small molecules (metabolites) present in a biological sample, such as blood serum or plasma, and offers a direct snapshot of an organism's physiological state. Its data, typically obtained by mass spectrometry or nuclear magnetic resonance, are high-dimensional and noisy, and often come from a limited number of samples, which makes them particularly hard to analyse.

Our projects cover the full analysis pipeline, from data preprocessing and feature selection to interpretable predictive models and the biological interpretation of the results. We currently work with serum and plasma data from two international cohorts, ADNI (Alzheimer's Disease Neuroimaging Initiative) and PPMI (Parkinson's Progression Markers Initiative), to identify metabolic profiles associated with Alzheimer's and Parkinson's disease, with a particular focus on the stability of the selected biomarkers across different algorithms, validation strategies and datasets.

Possible TFG/TFM projects

• Identification of metabolic biomarkers of Alzheimer's disease. Machine Learning applied to serum metabolomics data from ADNI to identify metabolic profiles associated with Alzheimer's disease and with the different stages of cognitive impairment. The project covers data preprocessing, feature selection, comparison of classification models and analysis of the relevance of the identified metabolites.
• Discovery of metabolic biomarkers of Parkinson's disease. Artificial Intelligence models to analyse plasma metabolomics data from PPMI and identify metabolites or metabolic signatures related to Parkinson's disease. The project studies how well different algorithms discriminate between clinical groups and analyses the stability and biological relevance of the selected biomarkers.
• Stability assessment of metabolic biomarkers. Analysis of the reproducibility of the metabolites selected by different algorithms, validation procedures and datasets, to determine which biomarkers remain consistently relevant and avoid conclusions that depend on a single sample or experimental setup.

Contact: David Bernal (dbernal@dia.uned.es)

This line develops techniques for analysing structural and functional magnetic resonance images to study the brain, including the localisation, characterisation and tracking of brain structures, and diagnostic support. Current technology makes a wide variety of non-invasive brain images available. Neuroimaging studies analyse the brain at different scales, seeking to identify the relationship between structure, function and behaviour. Among other topics, we focus on analysing both functional and structural MRI at different levels of abstraction.

In recent years we have taken part in several research projects on biomedical image analysis. These include project 018-ABEL-CM-2013, which developed computer vision techniques for characterising amorphous objects in 2D and 3D images, applied to the study of focal brain lesions. We have also worked on resting-state functional MRI (resting-state fMRI), developing clustering, classification and diagnostic support methods based on the study of the main functional brain networks.

Contact: Mariano Rincón (mrincon@dia.uned.es)

This line investigates Artificial Intelligence models capable of automatically analysing and assessing aesthetic aspects of images and visualisations, drawing on principles from psychology, neuroscience and design.

It builds on the G-GI3000/IDIK project, aimed at the automatic assessment of the landscape integration of infrastructures using visual information.

Contact: Mariano Rincón (mrincon@dia.uned.es)

Deep learning systems, including large language models and foundation models, perceive, generalise from data and generate content with remarkable power. However, they need large amounts of data, struggle to reason systematically and verifiably, are hard to interpret and do not explicitly incorporate expert domain knowledge. Symbolic systems, such as ontologies, knowledge graphs, logics and rules, offer the opposite: precise representation and inference, traceable reasoning and reusable knowledge, at the cost of being fragile with noisy data and costly to build and maintain.

This line studies and develops neuro-symbolic systems that combine the learning and perception capabilities of neural networks with the explicit reasoning of symbolic models, aiming at intelligent systems that are more accurate, interpretable and verifiable, and that need less data. The integration works in both directions: symbolic knowledge can guide, constrain or enrich neural learning, and neural techniques can build, complete or query symbolic structures. Applications include health, education, engineering, law, cultural heritage and other domains where expert knowledge can be formalised and explainability is required.

Possible TFM projects

• Knowledge-guided neural models. Integration of a domain ontology or knowledge graph with a neural model, and evaluation of the improvement over a purely neural approach.
• Question answering with knowledge graphs. Retrieval-augmented generation with knowledge graphs (GraphRAG), with analysis of the reliability and traceability of the answers.
• Semi-automatic construction of ontologies and knowledge graphs. Extraction of entities and relations from text with language models, with expert validation.
• Logical constraints in neural network training. Incorporation of domain rules, using approaches such as Logic Tensor Networks, to improve the consistency of predictions.
• Concept-based explanations. Explanation of the decisions of neural models in terms of the concepts and relations of an ontology.
• Evaluation of neuro-symbolic architectures. Experimental comparison and evaluation frameworks for reasoning, consistency and explainability, especially in low-data settings.

Neural computer vision models now recognise objects, people and simple actions with high accuracy, but identifying complex events in a continuous video stream remains an open problem. Events such as someone leaving an object out of place, carrying out a purposeful sequence of actions or departing from what is expected in an everyday situation are made up of objects, actions and spatial and temporal relations, and recognising them requires reasoning about context, domain knowledge and the history of the scene.

This line combines neural models of visual perception with ontologies, scene graphs and models of stereotyped situations to recognise composite events, out-of-place objects and deviations from the expected in real time. Neural components extract objects, poses, actions and spatial relations from the video, and symbolic components reason over them over time, producing explainable results with less training data and a computational cost suited to lightweight devices such as smart glasses. Applications include egocentric vision to assist blind and older people at home, natural-language queries about the video, and security or industrial monitoring. This research is carried out within the AI4BRAIN project.

Possible TFM projects

• Spatio-temporal scene graphs for event recognition. Building scene graphs from video and reasoning over them to recognise events defined in an ontology.
• Detection of out-of-place or missing objects. A symbolic memory of the scene state, fed by a neural object detector.
• Recognition of stereotyped situations in egocentric video. Domain scripts to recognise everyday activities and detect deviations from the expected sequence.
• Question answering about video. Integration of vision-language models with a symbolic reasoner to answer questions about what happened in the stream.
• Handling perceptual uncertainty. Probabilistic reasoning over events detected from noisy perception.
• Real-time processing on resource-limited devices. Optimisation of neuro-symbolic pipelines for edge devices such as egocentric cameras or smart glasses.

Early diagnosis of cognitive impairment (mild cognitive impairment, Alzheimer's disease and other dementias, and neurodegenerative and neurodevelopmental disorders) still relies largely on neuropsychological assessment. These tests are cheap, accessible and non-invasive, but their administration and interpretation depend on expert judgement and take up clinical time. They are also usually reduced to a final score, which leaves out the information contained in how the person performs the test.

This line applies Artificial Intelligence to neuropsychological tests and other non-invasive behavioural and cognitive markers, such as semantic and phonological fluency, spontaneous speech and discourse, graphical tests, and gestural and motor behaviour. The aim is to obtain assessment tools with greater discriminative power, able to detect the very earliest stages of decline and requiring fewer resources, so that they can be used in screening, primary care or remote assessment. Combining machine learning with knowledge representation allows models to incorporate expert neuropsychological criteria and explain their results, a key requirement for clinical acceptance.

Possible TFM projects

• Automatic analysis of verbal fluency. Analysis of semantic and phonological fluency tasks through clustering and switching, semantic distance between responses and the temporal dynamics of production, to discriminate between normal ageing, mild cognitive impairment and dementia.
• Analysis of speech and discourse. Acoustic, prosodic and linguistic features (pauses, disfluencies, lexical richness, syntactic complexity, coherence) extracted from picture descriptions, autobiographical narratives or spontaneous conversation.
• Digital biomarkers from the test-taking process. Analysis of how graphical or digital memory and executive tests are performed (stroke dynamics, timing, pauses, order of execution) and of its value compared with the classical score.
• Detection of patterns preceding clinical signs. Identification of early markers in people with subjective cognitive decline or risk factors, or through longitudinal follow-up.
• Hybrid models with expert knowledge. Models that combine machine learning with formalised neuropsychological criteria to produce explainable assessments.
• Multimodal models. Integration of graphical, verbal and gestural signals, including gestures, facial expression or gait captured on video.

This research line develops intelligent technologies to support assessment, intervention and follow-up in psychoeducational settings. Using Artificial Intelligence, Machine Learning and Knowledge Engineering, we design tools that help personalise learning, identify educational needs early and support the decisions of teachers and school counsellors.

Projects include intelligent recommender systems, educational assistants, predictive models of academic performance, psychoeducational assessment tools and support platforms for students with specific educational support needs.

Topics of interest

• Artificial Intelligence in Education.
• Learning analytics.
• Machine learning and educational data mining.
• Intelligent recommender systems.
• Knowledge Engineering and ontologies.
• Virtual Reality and immersive technologies.

This line develops Evolutionary Computation techniques for image processing, segmentation, detection and analysis. Projects explore how evolutionary algorithms can optimise computer vision processes in different application domains, especially medical imaging.

Previous projects include Computer Vision applications for the automatic analysis of medical images.

This line investigates the automatic design of models using Evolutionary Computation techniques. It addresses problems such as the design of electronic circuits, industrial models, hardware, mathematical functions, neural networks and other systems that can be optimised with evolutionary algorithms.

This line studies the foundations of Evolutionary Computation and the development of new evolutionary models, operators and mechanisms. Previous work includes research on selection mechanisms, evolution strategies and new approaches to improve the performance of evolutionary algorithms.