AI in Medical Image Computing

Increasing availability of medical imaging data with respect to
quantity, resolution and modalities, pose challenges to traditional
processing and analysis methods. Large cohort studies such as the
ADNI, Human Connectome Project, UK Biobank, Rhineland study
etc. collect rich data on 10-thousands of individuals. In order to
investigate etiology, progression, treatment, risk and preserving
factors of neurodegenerative diseases we need descriptive features,
obtained by scalable automatic processing methods. Traditional
approaches that depend on non-rigid atlas registration and
segmentation, however, are very slow (many hours up to a day for a
single image) and thus not efficient enough to handle big data or
provide quick results as needed in personalised medicine. We develop
fast AI methods for large multimodal datasets using deep learning that
can process images in minutes rather than hours or days. We will
demonstrate results of full brain segmentation and cortical
parcellation obtained in under 1 minute. We also use similar networks
successfully for fat segmentation in Dixon MRI. Furthermore, we
demonstrate that novel complex networks can be used to reconstruct
under-sampled images from K-space (raw MRI) data efficiently. Finally,
we will introduce advanced geometry-based features (e.g. shape and
lateral asymmetry analysis) that are sensitive to early disease
effects.