Smarter MRS with Physics and AI
Julian Merkofer defended his PhD thesis at the Department of Electrical Engineering on April 17.
Magnetic resonance spectroscopy (MRS) allows scientists to look beyond anatomy and into the chemistry of living tissue, such as the brain. But turning these signals into reliable insights has long been difficult. This PhD research of Julian Merkofer shows a key breakthrough: combining machine learning with explicit physical models of the MRS signal makes analysis significantly more robust, reliable, and informative than using AI alone. By blending data-driven methods with knowledge of how the signal is generated, the work offers a more trustworthy path toward real-world clinical use.
Unlike conventional MRI scans, which primarily reveal the structure of tissue, MRS captures its chemical composition. It detects metabolites鈥攕mall molecules involved in cellular processes鈥攑roviding a window into how tissue functions and how diseases develop.
However, this added depth comes at a cost. MRS data is notoriously difficult to interpret. Signals overlap, noise can obscure important details, and unwanted artifacts distort the measurements. These challenges have limited the widespread clinical use of MRS despite its potential.
Why traditional AI falls short
Machine learning has been widely explored as a solution to these challenges. In theory, AI can learn to interpret complex spectral patterns automatically. In practice, however, purely data-driven approaches often struggle.
One key issue is the gap between simulated training data and real-world (in-vivo) measurements. Models trained on clean, synthetic data can perform poorly when faced with messy, real signals. This lack of robustness makes them unreliable in clinical settings.
Combining physics and machine learning
The central idea of this thesis is simple but powerful: don鈥檛 let machine learning work alone. Instead, guide it using explicit physical models of how MRS signals are generated.
These hybrid approaches retain the strengths of classical physics-based modeling鈥攖ransparency and structure鈥攚hile benefiting from the flexibility of machine learning. The result is improved robustness, especially when moving from simulated data to real patient measurements.
Making results more trustworthy with uncertainty
In medical applications, knowing how certain a prediction is can be just as important as the prediction itself. This research of introduces a Bayesian framework that estimates uncertainty in a meaningful way.
Rather than producing a single 鈥渂est guess,鈥� the method provides a range of possible outcomes and shows how confident the model is. This makes the results more informative and better suited for clinical decision-making.
Cleaning up the signal: introducing WAND
Another key contribution is a method called WAND, designed to separate useful metabolic signals from unwanted components such as noise, baseline distortions, and artifacts.
By isolating the relevant information more effectively, WAND improves the accuracy of downstream analysis. This step is crucial for making MRS data easier to interpret and more reliable in practice.
Toward reliable and interpretable medical AI
Overall, the research demonstrates that model-based machine learning is a promising direction for MRS. It improves robustness, enhances interpretability, and makes the analysis better suited to real-world data.
Importantly, this work also speaks to a broader trend in medical AI: moving beyond raw prediction accuracy toward systems that are reliable, transparent, and trustworthy. In technically challenging fields like MRS, that shift may be essential for translating advanced methods into clinical impact.
Supported by the Spectralligence project (EUREKA IA Call, ITEA4 project 20209).
Title of PhD thesis: . Supervisors: Dr. Ruud van Sloun, and Prof. Massimo Mischi.