Research

The research in my group focuses on developing computational methods for analyzing organ-level medical images. Our aim is to significantly augment the capabilities of human experts, such as neuroscientists and clinicians, by analyzing images much faster, much more consistently, and in much more detail than they would otherwise be able to. This should ultimately enable new scientific insights into disease mechanisms, more efficient clinical trials, and improved treatment outcomes in individual patients.

The typical applications we address are automatic segmentation (delineating and measuring anatomical structures), registration (locally deforming images to match their content), subject-level prediction (automatic diagnosis or prognosis), and image reconstruction (turning physical measurements into images). Across all these applications, the focus of the group is on generative probabilistic models, as these excel at generalizing across scanners/protocols/diseases, at providing interpretable results that align intuitively with human reasoning, and at quantifying the uncertainty associated with the obtained results.

Publications

For my full list publication, see my Google scholar profile. Below I hightlight a few key publications addressing common themes in medical image computing.

Generalizability

Seldom addressed in academic benchmarks, but immediately surfacing in real-world deployments: the imaging zoo resulting from using different scanners, contrast mechanisms, contrast combinations, image resolutions, imaging modalities, patient populations, timing and number of follow-up scans in longitudinal imaging, ... The compositionality of generative models can deal with this without a combinatorial explosion of training data, often with no or only minimal changes to the actual software:

	An open-source tool for longitudinal whole-brain and white matter lesion segmentation. S. Cerri, D.N. Greve, A. Hoopes, H. Lundell, H.R. Siebner, M. Muehlau, K. Van Leemput. NeuroImage: Clinical, vol. 38, 103354, 2023
	A Contrast-Adaptive Method for Simultaneous Whole-Brain and Lesion Segmentation in Multiple Sclerosis. S. Cerri, O. Puonti, D.S. Meier, J. Wuerfel, M. Muehlau, H.R. Siebner, K. Van Leemput. NeuroImage, vol. 225, 117471, 2021
	A Modality-Adaptive Method for Segmenting Brain Tumors and Organs-at-Risk in Radiation Therapy Planning. M. Agn, P.M. af Rosenschold, O. Puonti, M.J. Lundemann, L. Mancini, A. Papadaki, S. Thust, J. Ashburner, I. Law, K. Van Leemput. Medical Image Analysis, vol. 54, pp. 220-237, 2019
	Fast and Sequence-Adaptive Whole-Brain Segmentation Using Parametric Bayesian Modeling. O. Puonti, J. E. Iglesias, K. Van Leemput. NeuroImage, vol. 143, pp. 235-249, 2016
	A computational atlas of the hippocampal formation using ex vivo, ultra-high resolution MRI: Application to adaptive segmentation of in vivo MRI. J.E. Iglesias, J.C. Augustinack, K. Nguyen, C.M. Player, A. Player, M. Wright, N. Roy, M.P. Frosch, A.C. McKee, L.L. Wald, B. Fischl, and K. Van Leemput. NeuroImage, vol. 115, pp. 117-137, 2015

Interpretability

Before automatic diagnostic systems can safely be adopted for widespread clinical use, they should not only predict well, but also explain their predictions in terms that are biologically meaningful. Generative models excel at this task, by mimicking the way humans look for relevant cause-effect relationships while ignoring irrelevant variations in the data:

A lightweight generative model for interpretable subject-level prediction. C. Mauri, S. Cerri, O. Puonti, M. Muehlau, K. Van Leemput. Medical Image Analysis, vol. 101, 103436, 2025

Uncertainty

The ability to provide end-users with reliable uncertainty estimates is a prerequisite in many clinical applications (e.g., when making life-changing treatment decisions). Generative models provide a principled way to compute accurate "error bars" using Markov chain Monte Carlo (MCMC) sampling:

	Fast Nonparametric Mutual-Information-based Registration and Uncertainty Estimation. M. Agn and K. Van Leemput. MICCAI2019 UNSURE workshop, Lecture Notes in Computer Science, vol. 11840, pp. 42-51, 2019
	Improved Inference in Bayesian Segmentation Using Monte Carlo Sampling: Application to Hippocampal Subfield Volumetry. J. E. Iglesias, M. R. Sabuncu, K. Van Leemput. Medical Image Analysis, vol. 17, no. 8, pp. 1181-1191, 2013

Analysis of well-known algorithms

We have analyzed some well-known algorithms over the years to show why they work (or sometimes don't!):

	Analysis of the Haufe transformation. (Don't use it!) Appendix D of: C. Mauri, S. Cerri, O. Puonti, M. Muehlau, K. Van Leemput. Medical Image Analysis, vol. 101, 103436, 2025
	A Cautionary Analysis of STAPLE Using Direct Inference of Segmentation Truth. K. Van Leemput and M.R. Sabuncu. MICCAI2014, Lecture Notes in Computer Science, vol. 8673, pp. 398-406, 2014
	N3 Bias Field Correction Explained as a Bayesian Modeling Method. C.T. Larsen, J.E. Iglesias, and K. Van Leemput. MICCAI2014 BAMBI Workshop, Lecture Notes in Computer Science, vol. 8677, pp. 1-12, 2014

Miscellaneous

I am one of the instigators of the well-known Multimodal Brain Tumor Segmentation Challenge (BRATS) series organized yearly in conjunction with MICCAI:

The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). B. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby, [...], M. Prastawa, M. Reyes, K. Van Leemput. IEEE Transactions on Medical Imaging, vol. 34, no. 10, pp. 1993-2024, 2015