Turku University of Applied Sciences

This study presents a robust and efficient client selection protocol designed to optimize the Federated Learning (FL) process for the Federated Tumor Segmentation Challenge (FeTS 2024). In the evolving landscape of FL, the judicious selection of collaborators emerges as a critical determinant for the success and efficiency of collective learning endeavors, particularly in domains requiring high precision. This work introduces a recommender engine framework based on non-negative matrix factorization (NNMF) and a hybrid aggregation approach that blends content-based and collaborative filtering. This method intelligently analyzes historical performance, expertise, and other relevant metrics to identify the most suitable collaborators. This approach not only addresses the cold start problem where new or inactive collaborators pose selection challenges due to limited data but also significantly improves the precision and efficiency of the FL process. Additionally, we propose harmonic similarity weight aggregation (HSimAgg) for adaptive aggregation of model parameters. We utilized a dataset comprising 1,251 multi-parametric magnetic resonance imaging (mpMRI) scans from individuals diagnosed with glioblastoma (GBM) for training purposes and an additional 219 mpMRI scans for external evaluations. Our federated tumor segmentation approach achieved dice scores of 0.7298, 0.7424, and 0.8218 for enhancing tumor (ET), tumor core (TC), and whole tumor (WT) segmentation tasks respectively on the external validation set. In conclusion, this research demonstrates that selecting collaborators with expertise aligned to specific tasks, like brain tumor segmentation, improves the effectiveness of FL networks.

Effective disease treatment often requires precise control of the release of the active pharmaceutical ingredient (API). In this work, we present a computational inverse design approach to determine the optimal drug composition that yields a target release profile. We assume that the drug release is governed by the Noyes-Whitney model, meaning that dissolution occurs at the surface of the drug. Our inverse design method is based on topology optimization. The method optimizes the drug composition based on the target release profile, considering the drug material parameters and the shape of the final drug. Our method is non-parametric and applicable to arbitrary drug shapes. The inverse design method is complemented by robust topology optimization, which accounts for the random drug material parameters. We use the stochastic reduced-order method (SROM) to propagate the uncertainty in the dissolution model. Unlike Monte Carlo methods, SROM requires fewer samples and improves computational performance. We apply our method to designing drugs with several target release profiles. The numerical results indicate that the release profiles of the designed drugs closely resemble the target profiles. The SROM-based drug designs exhibit less uncertainty in their release profiles, suggesting that our method is a convincing approach for uncertainty-aware drug design.

Generalization of imitation-learned navigation policies to environments unseen in training remains a major challenge. We address this by conducting the first large-scale study of how data quantity and data diversity affect real-world generalization in end-to-end, map-free visual navigation. Using a curated 4,565-hour crowd-sourced dataset collected across 161 locations in 35 countries, we train policies for point goal navigation and evaluate their closed-loop control performance on sidewalk robots operating in four countries, covering 125 km of autonomous driving. Our results show that large-scale training data enables zero-shot navigation in unknown environments, approaching the performance of policies trained with environment-specific demonstrations. Critically, we find that data diversity is far more important than data quantity. Doubling the number of geographical locations in a training set decreases navigation errors by similar to 15%, while performance benefit from adding data from existing locations saturates with very little data. We also observe that, with noisy crowd-sourced data, simple regression-based models outperform generative and sequence-based architectures.

As populations age globally, wearable health technologies offer promising solutions to support autonomy and well-being among older adults. This study explored the adoption of smart wearable systems, such as wristbands and chest sensors, for remote health monitoring among 352 older adults (aged 60–99) in Northern Portugal. Results showed that 74.4

Photovoltaic (PV) production grows rapidly in the Nordics, but literature on calculating long-term PV performance losses remains limited in these conditions, leading to inaccurate yield estimations and economic losses. To fill this knowledge gap, this contribution (1) shows typical PV performance characteristics, (2) obtains performance loss rates (PLRs) from multiple systems, and (3) assesses the best practices of PLR calculation under these conditions. For these objectives, we compared various methods for all steps of PLR calculation, including data filters, performance metrics, aggregation intervals, and statistical models for five systems (with 2.5-6 years of data) across Finland, from 60 degrees to 67 degrees N. Dark winters and snow coverage caused gaps and anomalies to the performance time series during wintertime, resulting in unrealistic PLRs. Excluding these anomalies with performance threshold filter improved the results. As the different methods can drastically affect the PLR value, an ensemble method consisting of an assessment of average PLR with over 700 calculation approaches was validated to be essential in identifying robust PLRs. Using this ensemble method, the system PLR estimates ranged from 0.41%/year to 2.65%/year. The data and codes are made publicly available, contributing to open access of PV datasets and allowing wider application of the PLR methodology.