Artificial Intelligence in Oncology

Supporting scientific research

Granted applications 2023

Alessandro Sbrizzi / Stefano Mandija

UMC Utrecht

In short, Sbrizzi's and Mandija's reserch entails the following:

Physics-informed Neural networks to standardize brain MRI: boosting AI applications in gliomas and meningiomas

In the field of magnetic Resonance (MR) imaging for brain tumors, a plethora of AI tools for tumor grading and contouring is theoretically available. However, these tools are trained on qualitative MRI images acquired with specific imaging settings, hence they work only when the same imaging settings are used. To overcome this limitation, we propose a technique which standardizes MRI data across systems, software, and imaging settings. Our novel MRI standardization technique will augment the MRI data training sets to enhance AI tools in the field of rare brain cancers such as meningioma and gliomas, where large standardized datasets are, at the moment, not available.

 

Publications

  • Quantitative mapping from conventional MRI using self-supervised physics-guided deep learning: applications to a large-scale, clinically heterogeneous dataset
    • Published in Arxiv (08-01-2026)
    • Click here to view the publication

 

Brief summary of progress / results April 2026

During the period April 2025 - April 2026, we developed and evaluated a physics-informed deep-learning framework for generating standardized quantitative MRI (qMRI) maps, with validation on clinical data from UMC Utrecht. The results were disseminated at major international imaging conferences and submitted to a leading medical imaging journal. Current work focuses on improving model robustness to better capture complex tissue behavior and on assessing whether standardized qMaps enable shortened MRI protocols by reducing the number of required sequences. In parallel, initial feasibility studies are underway to leverage standardized qMaps for deep-learning-based tumor detection and automatic segmentation.

 

Brief summary of progress / results April 2025

We are proud to announce that in the first 6 months of the project, we already managed to have a working version of the DL model for retrospective generation of standardized qMaps. We have tested this framework on a small research dataset of the UMC Utrecht and the results are presented as an oral presentation by our team member Jelmer van Lune at the Annual Meeting of the International Society for Magnetic Resonance in Medicine (ISMRM), in Honolulu, USA, May 2025