Rasmus Reinhold Paulsen
Department of Applied Mathematics and Computer Science
Richard Petersens Plads
Building 324 Room 110
2800 Kgs. Lyngby
3D surface acquisition medical image analysis Computational geometry geometric deep learning cardiovascular risk
My current research focus is on cardiovascular risk prediction from large 3D image databases. The close collaboration between my group at DTU and the heart center and the department for diagnostic radiology at Rigshospitalet enables us to train large deep learning models on truly unique image databases with associated demographics, biochemistry, cardiovascular outcome, and mortality data. Our goal is to uncover new or under-explored image biomarkers that can aid in the early detection of cardiovascular risks such as, for example, stroke or cardiac death. It is not trivial to use modern deep learning approaches on complex 3D data. We are exploring novel ways of representing 3D data including implicit surface descriptors and texture parameterizations, where our goal is to create compact and information-preserving models that can adapt to and harness the power of deep learning frameworks. One aspect of deep learning is the ability to reduce complex data into low dimensional latent spaces where facilitating separation of risk factors, patient groups, and other clinical end points. By exploiting and manipulating the geometry of the latent space, it is further possible to optimize networks to focus on the image features that are most significant for a given end point. We have previously demonstrated this on complex biomedical data and are extending this to the cardiovascular domain. Having effective 3D representations also has the potential for generative modelling of the human heart, where we can synthesize and visualize how known risk factors as smoking and alcohol affect the geometry and appearance of the heart leading to a better understanding of the impact of lifestyle on cardiac health.