Registration: A Swiss-Army Knife in Medical Image Analysis – Orcun Goksel

Although a well-studied topic, deformable image registration (DIR) is still far from being a solved problem, and it is arguably not yet fully exploited in many clinical contexts. DIR is a core problem in medical image analysis; but, unlike labeling decision problems such as classification and segmentation, registration is a problem class that involves stringent physical constraints. Although deep learning methods have made faster registration possible, the resulting models are often difficult to interpret compared to hand-crafted methods with explicit objectives and interpretable physical meaning.

This talk will explore a graph-labeling based registration approach as well as the utilization of registration in the context of weak-supervision of segmentation as anatomical priors and population models. These will be shown to be profitable for image datasets even with a lack of annotations, which is in turn is an important requirement for deep learning techniques. Solutions to other major registration challenges will also be demonstrated including efficient parametrization and computations of deformable fields, correct treatment of sliding interfaces, and estimates of uncertainty via consistency.

About Orcun Goksel