Generalizing deep Learning methods for volumetric medical image analysis

Abstract

The emergence of volumetric CT and MRI imaging technologies has dramatically improved clinical diagnostics and research, enabling visualization of body parts and organs in three dimensions. Deep learning, with its fundamental principles invented in the last century, has become a de facto standard for the automated processing of medical images, supporting clinicians in image interpretation and diagnosis. However, despite their widespread success, deep learning methods often achieve inferior results when applied in clinical practice compared to the training stage. This drop in performance is caused by the shifted properties of the images used during the deep learning models’ training and the images encountered later at the time of inference, combined with the models’ insufficient generalization capabilities. The shift in data properties to which the models fail to generalize may not be foreseen, and problematic image differences for the deep learning algorithms may be invisible to the human eye and not understandable by well-trained radiologists who can reliably diagnose patients’ conditions. In this thesis, four methods for volumetric medical imaging are presented that reliably generalize. It is researched in which areas and on which levels the generalization for volumetric medical images can be enabled and improved. The developed methods cover various fields of application, such as cardiac, abdominal, spinal, and brain volumetric medical imaging. Generalization was enabled by modeling acquisition processes for cardiac shape reconstruction, by effectively combining generalization and adaptation paradigms to overcome CT to MRI image intensity differences, by harnessing image registration in combination with loss-based modifications for generalizing segmentation of brain tumors across differently weighted MRI images, and by model parameter design modifications targeting the inner units of deep learning architecture to infer results from rotated or reflected input data reliably. All methods proved to work even for small-scale datasets with far less than one hundred samples, proving the efficiency of the methodological contributions as an alternative to following the trend of increasing dataset sizes and along with additional computational effort during training.