Image generated in biomedical diagnostic tasks are often acquired to monitor specific physiological, patho-physiological, or biophysical processes. These processes are underlying the given observables and the image features or biomarkers of interest. To this end, we are actively exploring the development of joint process and imaging models that, for example, describe tumor growth together with tumor-induced modifications of the image appearance, as well as efficient means for adapting these models to given image observations. In [Menze 2010, Menze 2016] we developed a multimodal tumor appearance model, and proposed a joint tumor-growth and -appearance model [Menze 2011] that can be adapted to multimodal observations using Bayesian methods [Lipkova 2020]. Related work includes the mult0-scale modeling of tumor dissemination [Praud 2019]. Recently, we demonstrated the use of deep learning techniques for inferring model parameters fast, both for PDE-based tumor growth models [Ezhov 2019], as well as for ODE-type signal model in blood perfusion [Ulas 2018, Rosa 2020], and in quantitative multi-parametric MR imaging [Gomez 2019].
The initial publication of this work [Menze 2010] was awarded the MICCAI Young Scientist Impact Award in 2015, which is awarded once a year by MICCAI society for the most influential publication of the past five years.
Reconstructing physical networks and graphs from 2D and 3D image data, for example, recorded from images displaying vascular networks, requires a special toolset of computational routines. Critical features for this computation are extreme ratios of fore- and background in vessel segmentation, the fine detail and small diameters of curvilinear tubular structures often at pixel resolution, and the global nature of constraints to be considered when tracing and linking edges and network branches. To this end we develop machine learning techniques that deal with these challenges when scaling computation to large real world data sets. Our prior work includes a first machine learning approach for centerline extraction [Schneider 2015], the extraction of networks under local shape constraints using integer programming [Rempfler 2015], and scaling the analysis to large 3D data sets using transfer learning [Todorov 2020]. Recent work includes the development of dedicated metrics to be used in segmenting fine curvilinear structures [Paetzold 2020].
A central publication of this work, [Rempfler 2015], has been awarded the ‘MICCAI best paper award’ in 2015.
Developing new computational algorithms for biomedical image quantification tasks requires openly available data sets. At the same time, the use of public data enables a comparison of competing algorithms and design concepts. To this end, the generation and curation of open data set is instrumental in a bi-directional translational effort: on the one hand it formalizes biomedical image processing task and makes well-prepared data set available to computing specialists, on the other hand it encourages a backward translation of top performing algorithms into a biomedical image processing pipeline. I pioneered this work by initiating the Multimodal Brain Tumor Image Segmentation Challenge (BRATS) that, since its first edition in 2012, became one of the landmark challenges in biomedical image computing [Menze 2015, Bakas 2019]. We repeated this effort on other challenges on organ and lesion segmentation
(VISCERAL, ISLES, LITS, Decathlon, QUBIQ) [Maier 2015, Toro 2016, Christ 219, Sekuboyina 2020], and contributed to coordinating efforts for establishing rules for best practice [Maier-Hein 2019].
In particular the BRATS paper demonstrates the impact of this work: The study is the most cited paper of TMI for five years in a row, with about 3000 citations at present. It led to a significant number of MICCAI paper on the topic, opening new directions serving as a landmark application for biomedical image segmentation. It is also shaping industry standards, for example, by being one of Intel’s reference data sets for performance benchmarking and optimization of CNN primitives and public topologies.