Andrei Kazlouski

We present a comprehensive validation methodology for federated learning models, evaluated on an extensive collection of publicly available datasets from peer-reviewed medical publications. Our main goal is to answer the question: “When is it truly beneficial for a hospital to participate in federated learning training?”. We further extend this methodology to a practical prostate cancer use case. Using three nodes located in Finland and Belgium, we established a fully functional, non-simulated (i.e., real) federated learning setup operating on private medical data.

Juha Pajula, Jaakko Lähteenmäki

VTT has developed a dedicated assessment tool which helps technology developers and EHDS providers to test how well the targeted technology is aligned with EHDS and related regulation. At the initial phase the tool aims to support assessment of PHASE IV AI project outputs, which are intended to be used in the context of European Health Data Space (EHDS) infrastructure. The checklist is applicable to technical outputs, like tools, algorithms, models and platforms. The assessment process mainly concerns compliance from the regulatory perspective but also helps to assess if the functionality of the output is relevant for EHDS. Currently three output categories are considered based on EHDS terminology: Secure Processing Environment resource, resource platform, and health data intermediation platform. The regulatory compliance components of the tool are based on the current EHDS and GDPR regulations. As several details of EHDS regulation are still open and subject to implementing acts the tool will need new reviews after all acts are implemented in EU and national level. Where current regulation does not provide sufficient details, the checklist items have been derived based on relevant sector practices, existing sources such as TEHDAS2 project deliverables and expert knowledge.

Ayşen Değerli, Gaurang Sharma, Harri Pölönen, and Juha Pajula (Presenter: Tunc Asuroglu)

We trained AI models to generate synthetic 3D magnetic resonance imaging (MRI) data of human brains with stroke lesions. These models were trained with ATLAS R2.0 (Anatomical Tracings of Lesions After Stroke) dataset with 655 images from 33 subjects. Leaving out 93 images from 6 subjects for testing. After literature search and experiments, we chose latent diffusion-based AI model and further developed two different variations of it. The image quality and generated brain anatomy in the model output is close to realistic but there were problems with image uniqueness and diversity.
The utility of synthetic data was tested with training task where we trained deep learning segmentation models to segment ischemic stroke lesions from 3D MRI data. The models were trained both synthetic and ATLAS R2.0 dataset for comparison. The selected AI models were 3D versions of the well-known UNet models, where we tested UNet and UNetR models first one having a basic U-structured and second a transformer-induced UNet. Models were tested with the test set, which was not included in training dataset of synthetic model generator nor Unets trained with real data. Our results showed that compared to the transformer-induced UNetR model, UNet models achieved better performance with the sliding window inference being also with faster to train. The models trained with synthetic data were having fair performance but lesser than models trained with real data.

Sami Pietilä*, Teemu Herttua, Joni Obradovic, Axel Bäcklund, Kaisa Jokela, Khaoula El Mekkaoui, Irfan Khan, Heikki Lassila, Elina Kontio, Eero Lehtonen, Jussi Salmi

We describe our method for FL, the data that we have used in training, and the training results. We don’t have piloting results yet. The datasets are open access, but they have a variety of modalities and image resolution and ground truth segmentation masks can have problematic features with regard to using them as training material. The treatment processes introduce restrictions to how the segmentation should be done, because certain modalities are preferred in the diagnosis stage. Initial results of the segmentation are not satisfactory, considerable preprocessing is needed to improve the results and finally our methods were able to produce segmentation models with good error metrics.

Anna Oliveras, Roger Marí, Rafael Redondo, Oriol Guardià-Olivella, Ana Tost, Bhalaji Nagarajan, Carolina Migliorelli, Vicent Ribas, Petia Radeva

This work introduces a new latent diffusion model to generate high-quality 3D chest CT scans conditioned on 3D anatomical masks. The method synthesizes volumetric images of size 256x256x256 at 1mm isotropic resolution using a single mid-range GPU, significantly lowering the computational cost compared to existing approaches. The conditioning masks delineate lung and nodule regions, enabling precise control over the output anatomical features. Experimental results demonstrate that conditioning solely on nodule masks leads to anatomically incorrect outputs, highlighting the importance of incorporating global lung structure for accurate conditional synthesis. The proposed approach supports the generation of diverse CT volumes with and without lung nodules of varying attributes, providing a valuable tool for training AI models or healthcare professionals. Available at: https://doi.org/10.1038/s41598-026-51634-4

Congting Lai, Christos Chatzichristos, Maarten De Vos

Synthetic data is typically evaluated using aggregate utility or fidelity metrics computed over the full dataset. However, real-world data distributions are highly heterogeneous and imbalanced, where dominant easy samples can obscure substantial failures on difficult, rare, or underrepresented regions of the data distribution. In this work, we move beyond average evaluations and adopt a subset-based perspective for assessing synthetic data. We show that strong global downstream performance, such as high AUC, does not necessarily imply faithful preservation of performance behavior across the full subset spectrum.To systematically evaluate performance across varying levels of subgroup difficulty, we introduce AUSC (Area Under the Subset Curve),a metric that characterizes predictive performance along a worst-to-global subset trajectory. AUSC captures how performance degrades across increasingly challenging subpopulations. We further extend this to Worst-X-AUSC, enabling targeted evaluation of tail-risk regions. We validate this metric through distributional fidelity checks and interpretable subgroup discovery, demonstrating its ability to reveal mearningful subset failures that are not captured by aggregate evaluations.

Diogo Martins, Pedro Sousa, Antonio Cardoso, Daniela Ferreira-Santos, Helder P. Oliveira, Tania Pereira

Lung cancer remains one of the leading causes of cancer-related mortality worldwide, with early detection playing a critical role in improving patient outcomes. Computed tomography (CT) imaging is a cornerstone of lung cancer diagnosis; however, the development of robust artificial intelligence (AI) models is often constrained by the scarcity of high-quality, annotated medical imaging data and concerns about patient privacy.

Rui Jorge Ramos, Adebayo Abayomi-Alli, Pedro Fernandes Sousa, Gonçalo Gonçalves, Artur Rocha

Segmenting 3D lung CT scans is crucial for pulmonary diagnostics, but training robust deep learning models for this task requires massive, diverse datasets. In healthcare, this immediately hits a bottleneck: patient data privacy. Privacy-Preserving Federated Learning (PPFL) offers a practical workaround, allowing institutions to train models collaboratively without moving or exposing sensitive data.
In this ongoing work, we are implementing a UNet-based 3D lung CT segmentation model with PPFL mechanisms. Our goal is to understand exactly how these privacy layers will impact segmentation performance and training efficiency. To do this, we plan to evaluate the model across three different setups: a standard local training baseline, an in-house federated environment using distributed virtual machines, and a future federated deployment on the TuneInsight platform for multi-partner collaboration. By comparing these three scenarios, we aim to map out the trade-offs between segmentation accuracy, privacy guarantees, and infrastructure overhead. Ultimately, we expect our findings to provide a realistic assessment of the feasibility of deploying PPFL for heavy 3D medical imaging tasks across institutions.

Ceren Yıldırım, Ali Şah Özcan, Tolun Tosun*, Erkay Savaş

Modern machine learning (ML) increasingly relies on sensitive data, especially in domains such as healthcare. However, using such data (e.g., confidential patient records) in external computing environments raises important privacy, security, and regulatory concerns. Fully Homomorphic Encryption (FHE) enables computation directly on encrypted data, allowing servers to process inputs without seeing the underlying plaintext. The main challenge is performance: FHE operations are computationally expensive, making acceleration essential for practical privacy-preserving ML.
We introduce HEonGPU, a high-performance library that accelerates FHE operations on Graphics Processing Units (GPUs). By utilizing GPU parallel processing capabilities, HEonGPU reduces the computational burden of FHE, making it viable for time-sensitive tasks. It supports various FHE schemes (BFV, CKKS, TFHE) and features a user-friendly C++ interface with an optimized multi-stream architecture for efficient memory management. We demonstrate HEonGPU’s effectiveness as an acceleration layer in privacy-preserving machine learning pipelines, enhancing operational efficiency while maintaining data privacy.

Akarsh K Nair, Arifur Rahman, Nickolas Shopland, Andy Burton and David Brown*

This work presents a privacy-preserving AI-driven cancer screening risk assessment tool for distributed electronic health record (EHR) environments. The platform combines federated learning, calibrated neural network prediction, and explainable AI to support interpretable clinical risk screening without centralising sensitive patient data. The system enables real-time cancer risk estimation using demographic, respiratory, cardiovascular, diabetic, and lifestyle-related indicators through an interactive interface. Federated IID, federated non-IID, and centralised learning configurations were evaluated to assess predictive robustness under heterogeneous healthcare settings.
Integrated Gradients–based explainability was incorporated to improve transparency and support interpretation of model predictions. The proposed framework demonstrates the potential of federated and explainable AI for privacy-aware clinical decision support and scalable cancer screening applications.

Mariya Georgieva*, Thierry Bossy*

This poster illustrates a real-world healthcare scenario in which multiple clinical institutions collaborate on AI-driven research without ever sharing sensitive patient data. Starting from raw, heterogeneous clinical data across sites, we walk through the full pipeline: data harmonization and interoperability using the OMOP common data model, through a privacy-preserving federated learning framework that manages coordination, model training, and governance across institutions, all the way to clinically relevant outcomes in oncology.
We highlight two concrete use cases — prostate cancer and lung cancer — spanning diverse data modalities including structured clinical records and complex 3D MRI imaging. Underpinning the framework are advanced privacy technologies including Federated Learning and Fully Homomorphic Encryption, ensuring that sensitive data never leaves its source institution. The poster demonstrates how these components come together to enable scalable and trustworthy AI for real clinical research.