Strong ties between KU Leuven, the University of Helsinki and Aalto University have been elevated to the international innovation stage thanks to the Exploratory Seed Funding. As part of this collaboration, the group is inviting you to the first workshop on trustworthy AI for medical image analysis and computer vision.
Date: Tuesday, May 27th, from 9:00-16:00 EEST
The venue has changed.
Updated location: The workshop will take place in Scandic Grand Central Hotel (Vilhonkatu 13, 00100 Helsinki, next to the Central railway station) and remotely via ZOOM.
Who can join?
The workshop is open for researchers with different backgrounds interested in trustworthy AI and medical image analysis.
Registration
Please register HERE at the latest by May 16, 2025. The space for in-person participation is limited. To secure your place in person, we recommend you register as soon as possible.
Note: registration is mandatory.
Have any questions?
Contact Katarina (katarina.sladakovic@helsinki.fi) should you have any questions or technical difficulties.
9:15-9:30 Registration
9:30-9:40 Workshop opening
9:40-10:40 Session 1
9:40-10:10 Modeling annotation noise in medical image segmentation by Aleksei Tiulpin, University of Oulu
10:10-10:40 Uncertainty Calibration via Proper Scores for Trustworthy AI by Sebastian Gruber, KU Leuven
10:40-11:00 Coffee break
11:00-12:00 Session 2
11:00-11:30 Medical Image Segmentation with SAM-generated Annotations by Juho Kannala, Aalto University
11:30-12:00 Enhanced federation in developing clinically useful algorithms for the infant EEG analyses by Sampsa Vanhatalo, University of Helsinki
12:00-13:30 Lunch
13:30-14:30 Session 3
13:30-14:00 Privacy risks in deep learning and how to prevent them by Antti Honkela, University of Helsinki
14:00-14:30 Large-scale MEG and EEG data aggregation: challenges and opportunities by Lauri Parkkonen, Aalto University
14:30-14:50 Coffee break
14:50-15:50 Session 4
14:50-15:20 Towards generalizable EEG models via federated and self-supervised pre-training by Tim Hermans, KU Leuven
15:20-15:50 Private Large Language Model Adaptations for Supporting Medical Triage by Franziska Boenisch, CISPA
15:50-15:55 Closure of the workshop
15:55-16:00 Break
16:00-17:00 Brainstorming and wrap-up (internal discussion, invitation only)
Modeling annotation noise in medical image segmentation by Aleksei Tiulpin, University of Oulu
Abstract: Prevention is the next frontier of modern medicine, and machine learning (ML) will undeniably play a pivotal role in making it a reality. Medical imaging enables objective tissue quantification, robust diagnostics, and treatment monitoring. Unlocking the power of these high-dimensional data often depends on accurate image segmentation, but the inherent uncertainty in medical scans leads to noisy annotations. The current approach to handle it is through large data collection and large-capacity models. The present talk features my lab’s work on modeling uncertainty in medical image segmentation (MIS) using prior knowledge communicated to models through soft labels. We show that by treating MIS as a structured regression problem, despite the current trends, high-capacity models may not be necessary. Extensive and rigorous benchmarking of our methods on brain tumor segmentation and retinal vessel segmentation reveals that a simple UNet with properly modeled annotation noise can achieve a lot better performance than methods published in the last 5 years in these domains.
Uncertainty Calibration via Proper Scores for Trustworthy AI by Sebastian Gruber, KU Leuven
Abstract: This talk highlights recent advances in understanding and improving the trustworthiness of classifiers in artificial intelligence (AI) via uncertainty calibration. The calibration of a classifier quantifies the reliability of its probability predictions and can be further improved via recalibration methods. However, estimating this calibration in practice is usually biased and inconsistent. Here, we introduce the broad framework of proper calibration errors with a general estimator. Our framework relates calibration to proper scores, which are an axiomatic class of loss functions for probabilistic predictions. This relationship is used for a more reliable recalibration of classifiers, which we demonstrate with experiments in common image classification settings. We theoretically and empirically demonstrate the shortcomings of commonly used estimators compared to our approach. In summary, our contribution offers principled tools for building more trustworthy AI systems.
Medical Image Segmentation with SAM-generated Annotations by Juho Kannala, Aalto University
Abstract: The field of medical image segmentation is hindered by the scarcity of large, publicly available annotated datasets. Not all datasets are made public for privacy reasons, and creating annotations for a large dataset is time-consuming and expensive, as it requires specialized expertise to accurately identify regions of interest (ROIs) within the images. To address these challenges, we evaluate the performance of the Segment Anything Model (SAM) as an annotation tool for medical data by using it to produce so-called "pseudo labels" on the Medical Segmentation Decathlon (MSD) computed tomography (CT) tasks. The pseudo labels are then used in place of ground truth labels to train a UNet model in a weakly-supervised manner. We experiment with different prompt types on SAM and find that the bounding box prompt is a simple yet effective method for generating pseudo labels. This method allows us to develop a weakly-supervised model that performs comparably to a fully supervised model.
Towards generalizable EEG models via federated and self-supervised pre-training by Tim Hermans, KU Leuven
Abstract: This presentation explores the potential of federated learning for developing EEG models that generalize across decentralized and heterogeneous datasets. The session will begin with a brief overview of our group's expertise and recent advances in AI-based EEG interpretation. Building on this foundation, the second part explores how we believe federated learning frameworks—which provide access to large amounts of distributed, unlabeled data—can be leveraged to train more robust models. A proof-of-concept study is presented in which the added value of self-supervised and federated pre-training strategies are evaluated in a simulated decentralized setting. Using a neonatal EEG dataset, models trained solely with conventional supervised learning on a small labeled subset are compared to models that include self-supervised pre-training on larger unlabeled datasets, both in centralized and federated configurations. Preliminary results suggest that self-supervised pre-training can significantly improve generalization to unseen data, highlighting a key advantage of federated learning. Additionally, challenges that federated learning presents in this context are discussed.
Private Large Language Model Adaptations for Supporting Medical Triage by Franziska Boenisch, CISPA
Abstract: As language models (LLMs) underpin various sensitive applications, such as medical ones, preserving privacy of their training data is crucial for their trustworthy deployment. This talk will focus on the privacy of LLM adaptation data. We will see how easily sensitive data can leak from the adaptations, putting privacy in risk. We will then dive into designing protection methods, focusing on how we can obtain privacy guarantees for adaptation data, in particular for prompts. This research is part of our ILLUMINATION research project that aims at deploying privately adapted LLMs to support medical triage in the emergency unit.