Preview

PDF, English - main document Download (11MB) | Terms of use

Abstract

The emergence of new data- and algorithm-driven analysis methods is revolutionizing many areas of research and enabling solutions to problems that were previously considered intractable. But does this also apply to medicine?

Improving diagnosis, fine characterization of patients for personalized medicine, monitoring disease progression and its prognosis, or predicting the outcomes of various therapies are just some of the areas that could potentially benefit from such new analysis techniques based on Deep Learning, which has enabled major advances in computer vision and related fields. Such algorithms now enable machines to better understand and interpret visual image data, which on the one hand offers promising perspectives for medical image processing, but on the other hand also poses challenges. As such, large amounts of annotated data are needed to prevent overfitting and to generate robust, generalizing and reliable models. Multicenter imaging studies could greatly improve the availability of such data and also enhance heterogeneity by obtaining training data from various sites. However, sharing data across multiple sites has proven to be difficult due to the high level of data protection associated with medical records and technical challenges such as interoperability. Consequently, this thesis attempts to avoid the necessity of data exchange by following a different approach: "Let’s share the algorithms, not the data!"

The objective and central research question here is whether it is possible to shift the evaluation and training of modern image analysis methods to the clinical data owners and how this can be accomplished. Although this approach helps to circumvent the data export and the associated data protection challenges, the participating partners must still be enabled to execute the algorithms from a technical and or- ganizational perspective. For this purpose, this dissertation investigated concepts, the development and establishment of a decentralized infrastructure for clinical medical image analysis that enables standardized data access and uniform execution of analysis methods for joint data analysis in the context of multicenter imaging studies. The technical realization of this infrastructure was achieved by developing a software framework called Kaapana, which enables the building of imaging platforms. The resulting software can be hosted on dedicated servers within the clinical IT environ- ment to be operated isolated from any external connectivity and to be interconnected with other local clinical systems. By leveraging modern cloud technologies such as containers and Kubernetes, the local deployment provides a private cloud for image processing that can be accessed from any locally connected workstations via the web browser. Standardized linkage to the clinical PACS via DICOM and the integration of a research imaging archive enables redundant data management and consistent data access for the execution of analysis methods. Dashboards enable efficient data exploration and filtering by visually presenting DICOM metadata of the platform’s data and allowing it to be selected via search queries. A uniform execution environment for data processing allows algorithms to be applied to such selected data and to be uniformly packaged and shared with partners. High-performance server hardware including Graphics Processing Units enables a variety of analysis techniques such as Deep Learning-based model inference, as well as the training of new models to be shared with partners. Within the infrastructure, the standardized and widely adopted DICOM format has been prioritized so that also many analysis results can be provided in a standard-compliant way. Using formats such as DICOM SEG, SR or Encapsu- lated PDF, data annotations become compatible with clinical workflows and IT systems. With support of the Radiological Cooperative Network and the German Cancer Consor- tium, the resulting infrastructure could be deployed and evaluated within two German research consortia. To this end, all 36 German university hospitals have commissioned their own server on which the platform has been installed and tested. This involved evaluating the commissioning, integration, operation and maintenance of such a decentralized network, as well as various use cases designed to cover the typical tasks of such a system. As a result, these use cases and the corresponding varying demands could be realized with the developed concept and the implemented framework. A flexible extension mechanism also allows the integration of additional services such as137 the MITK workbench or processing algorithms such as the nnUNet into the framework. Furthermore, first analysis pipelines developed by external partners could also be integrated and delivered to the partners already.

Within the scope of this work, clinics were enabled to apply up-to-date research methods to their own data through the development, distribution and support of the developed infrastructure. Since the research consortia, which already include all German university hospitals, have only just started their activities, there are great opportunities to make the high-quality data from the partner sites accessible and usable for research in the future. Because of Kaapana’s open source code, its architecture based on common industry standards, and its already broad deployment, this framework could also serve as a foundation for areas other than medical imaging, and thus offer the potential for tighter data integration for clinical computing in general.