Disputation Michael Baumgartner Generalised Medical Object Detection via Self-Configuring Method Design

The increasing number of volumetric medical image acquisitions is resulting in a growing workload for clinicians, ultimately fueling the need for automated diagnosis through Computer-Aided Diagnosis (CAD) tools. Clinical decision making often relies on diagnostic tasks which can be directly translated into the localization and classification of critical pathologies such as vessel occlusions, lung nodules, aneurysms, and tumors. Object detection methods can learn to identify such pathologies in an end-to-end fashion, providing great utility by directly solving diagnostic tasks. However, the adoption of these methods in the medical domain is hindered by limited experience and the complex configuration of application-specific parameters. This thesis first presents three case studies where detection methods were manually configured: (1) for the identification of mediastinal lesions, (2) for the detection of vessel occlusions, and (3) an exploratory study to assess the feasibility of Detection Transformer (DETR) models for volumetric images. Subsequently, the first self-configuring volumetric detection method is introduced, named nnDetection. This method was developed and evaluated on an unprecedented number of volumetric detection tasks and was used as part of multiple winning entries in international competitions. This thesis builds the foundation of future research in this domain by leading the shift towards the development of generalizable medical object detection methods.