The surge of supervised learning methods for segmentation lately has underscored the critical role of label quality in predicting performance. This issue is prevalent in the domain of medical imaging, where high annotation costs and inter-observer variability pose significant challenges. Acquiring labels commonly involves multiple experts providing their interpretations of the “true” segmentation labels, each influenced by their individual biases. The blind acceptance of these noisy labels as the ground truth restricts the potential effectiveness of segmentation algorithms. Here, we apply coupled convolutional neural network approaches to a small-sized real-world dataset of bovine cumulus oocyte complexes. This is the first time these methods have been applied to a real-world annotation medical dataset, since they were previously tested only on artificially generated labels of medical and non-medical datasets. This dataset is crucial for healthy embryo development. Its application revealed an important challenge: the inability to effectively learn distinct confusion matrices for each expert due to large areas of agreement. In response, we propose a novel method that focuses on areas of high uncertainty. This approach allows us to understand the individual characteristics better, extract their behavior, and use this insight to create a more sophisticated ground truth using maximum likelihood. These findings contribute to the ongoing discussion of leveraging machine learning algorithms for medical image segmentation, particularly in scenarios involving multiple human annotators.

Cumulus expansion is an important indicator of oocyte maturation and has been suggested to be indicative of greater oocyte developmental capacity. Although multiple methods have been described to assess cumulus expansion, none of them is considered a gold standard. Additionally, these methods are subjective and time-consuming. In this manuscript, the reliability of three cumulus expansion measurement methods was assessed, and a deep learning model was created to automatically perform the measurement. Cumulus expansion of 232 cumulus-oocyte complexes was evaluated by three independent observers using three methods: (1) measurement of the cumulus area, (2) measurement of three distances between the zona pellucida and outer cumulus, and (3) scoring cumulus expansion on a 5-point Likert scale. The reliability of the methods was calculated in terms of intraclass-correlation coefficients (ICC) for both inter- and intra-observer agreements. The area method resulted in the best overall inter-observer agreement with an ICC of 0.89 versus 0.54 and 0.30 for the 3-distance and scoring methods, respectively. Therefore, the area method served as the base to create a deep learning model, AI-xpansion, which reaches a human-level performance in terms of average rank, bias and variance. To evaluate the accuracy of the methods, the results of cumulus expansion calculations were linked to embryonic development. Cumulus expansion had increased significantly in oocytes that achieved successful embryo development when measured by AI-xpansion, the area- or 3-distance method, while this was not the case for the scoring method. Measuring the area is the most reliable method to manually evaluate cumulus expansion, whilst deep learning automatically performs the calculation with human-level precision and high accuracy and could therefore be a valuable prospective tool for embryologists.

In this study, we present a comprehensive pipeline to train and compare a broad spectrum of machine learning and deep learning brain clocks, integrating diverse preprocessing strategies and correction terms. Our analysis also includes established methodologies which have shown success in prior UK Biobank-related studies. For our analysis we used T1-weighted MRI scans and processed de novo all images via FastSurfer, transforming them into a conformed space for deep learning and extracting image-derived phenotypes for our machine learning approaches. We rigorously evaluated these approaches both as robust age predictors for healthy individuals and as potential biomarkers for various neurodegenerative conditions, leveraging data from the UK Biobank, ADNI, and NACC datasets. To this end we designed a statistical framework to assess age prediction performance, the robustness of the prediction across cohort variability (database, machine type and ethnicity) and its potential as a biomarker for neurodegenerative conditions. Results demonstrate that highly accurate brain age models, typically utilising penalised linear machine learning models adjusted with Zhang's methodology, with mean absolute errors under 1 year in external validation, can be achieved while maintaining consistent prediction performance across different age brackets and subgroups (e.g., ethnicity and MRI machine/manufacturer). Additionally, these models show strong potential as biomarkers for neurodegenerative conditions, such as dementia, where brain age prediction achieved an AUROC of up to 0.90 in distinguishing healthy individuals from those with dementia.

Wildfires are becoming more frequent and severe in many fire-prone regions, with disproportionate impacts on carbon emissions, ecosystems, and society. However, existing fire and Earth system models still struggle to represent the highly localized and stochastic nature of extreme fire ignitions and to quantify their short-term impacts at fine spatial scales.In this work, we develop a data-driven framework for next day active fire forecasting at sub-kilometre resolution by combining reanalysis meteorology with satellite fire observations. Our approach builds on recent advances in spatio-temporal deep learning from the AI community, in particular Earth system transformers and denoising diffusion probabilistic models. We use multi-year ERA5 meteorological fields together with static variables (topography, land cover, fuel proxies) as training data, and VIIRS active fire detections and pixel-level brightness/fire radiative power as targets.The model consists of a deterministic spatio-temporal backbone that encodes the joint evolution of weather and surface conditions, coupled to a diffusion-based probabilistic head that predicts the distribution of future ignition locations and associated fire intensity. This design allows us to explicitly represent uncertainty in rare, extreme events while retaining high spatial resolution. We evaluate the system on multiple fire-prone regions and held-out seasons containing documented extreme fire episodes. Preliminary results show improved skill in localizing ignitions and capturing extreme-tail intensity compared to baseline statistical and convolutional models, particularly in top-k precision metrics relevant for operational targeting.We plan to couple the predicted intensity fields with standard emission factors to estimate event-scale CO₂ emissions and explore the relative importance of meteorological and surface drivers using feature attribution techniques using causality discovery methods. Our findings illustrate the potential of modern probabilistic deep learning to bridge between high-resolution fire observations and Earth system applications, and to support the assessment and management of future extreme fires.

Droughts are hydroclimatic anomalies driven by precipitation deficits and increased evapotranspiration, posing an escalating threat under a warming Mediterranean climate. Assessing drought risk remains challenging due to the complex interactions between biophysical conditions and human systems, as well as limitations in impact reporting. Moreover, drought impacts are highly heterogeneous across sectors, as different types of drought affect socio-environmental systems differently. In this context, we develop a spatio-temporal deep learning framework to predict sector-specific drought impacts and identify the environmental and climatic drivers of these impacts.We combine two primary data sources: the European Drought Impact Database (EDID), which contains above 13,000 georeferenced drought impact reports spanning 1970 to 2023 and aggregated into four sectors (agriculture, ecosystem, energy, and socio-economic); and a set of physical drivers, including precipitation, temperature, drought indices, vegetation indices, and population density, derived from various sources for the period 2001–2021.The prediction task is formulated as a spatio-temporal segmentation problem using a 3D U-Net architecture to capture dependencies in climate and environmental conditions over a one-year period. The preprocessing workflow harmonizes all variables to a spatial resolution of 0.25° and an 8-day time step. Seasonally varying predictors are transformed into anomalies, and all variables are normalized. Input samples are arranged as tensors with shape 36×48×16×16 (C×T×H×W), representing one year of conditions, while the target consists of a binary impact map (1×1×16×16) corresponding to the subsequent 8-day period. The training dataset is balanced through equal sampling of impact and no-impact cases. Consequently, the model learns to use one year of spatio-temporal context to predict drought-affected areas at the next time step.Initial results for the agricultural sector indicate that traditional drought indices have limited predictive skill for drought impacts. A baseline evaluation of the Standardized Precipitation-Evapotranspiration Index (SPEI) across multiple thresholds shows that the 1-month SPEI achieves a PR-AUC of 0.13 and an ROC-AUC of 0.32 for the impact class over the 2018-2020 test period, with similarly low performance for the 3-, 6-, and 12-month variants. In contrast, preliminary model experiments demonstrate a substantial improvement over the baseline, achieving an F1 score of 0.43, a PR-AUC of 0.41, and a ROC-AUC of 0.71, despite remaining limitations in predictive performance.These limitations are primarily attributed to noise and spatial uncertainty in the ground-truth labels, as EDID impacts are reported at coarse administrative units (NUTS3) and uniformly assigned to all grid cells within each region, constraining pixel-level learning. In addition, drought impacts are influenced by large-scale atmospheric circulation patterns and remote climate teleconnections (e.g., ENSO and NAO) that are not explicitly represented in the current feature set. Future work will address these limitations by incorporating large-scale circulation and teleconnection indicators, developing strategies to mitigate label noise, and extending the modelling framework to additional sectors. Once predictive performance is optimized, explainable AI methods based on Integrated Gradients will be applied to identify the most influential climatic and environmental drivers, enabling sector-specific interpretation of drought impact mechanisms and their temporal dynamics.