Annan Halvtidskontroll Fabian Sinzinger

Huvudhandledare 

Docent Rodrigo Moreno, Division of Biomedical Imaging, Kungliga tekniska högskolan 

Bihandledare

Docent Joana Pereira, Institutionen för klinisk neurovetenskap, Karolinska Institutet 

Professor Örjan Smedby, Division of Biomedical Imaging, Kungliga tekniska högskolan 

Halvtidsnämnd 

Professor Michael Felsberg, Institutionen för systemteknik, Linköping universitet

Professor Joakim Lindblad,  Institutionen för informationsteknologi, Uppsala universitet

Professor Tomas Bjerner, Institutionen för hälsa, medicin och vård, Linköping universitet

Kort beskrivning 

In recent years, general deep learning-based methods have been vastly successful for medical image processing problems. This novel, data-driven methodology introduces high demands on the availability of big datasets for model training, which is an issue in many medical image processing tasks. Geometric deep learning (GDL) is a family of deep learning methods that utilise mathematical concepts of structures, symmetries, and in- or equivariances. Those intrinsic geometric properties of the data and the underlying domains are used here to build problem-specific models. This PhD project demonstrates how GDL can be applied to concrete medical image processing problems. Moreover, the project shows that GDL is a viable alternative for problems with small training datasets. The problems addressed in this thesis include trabecular bone stiffness estimation from micro-CT data, CT-based lung cancer survival rate prediction, tractography extracted from diffusion MRI data, and structural brain connectivity.