Addressing data heterogeneity in image-based computer-aided detection and diagnosis methods

Abstract

The acquisition of medical imaging data is inevitable for screening, diagnosis, planning of surgery or therapy, or monitoring of diseases. In clinical practice, the data is assessed by medical experts, which can be a very time-consuming task. Hence, for decades a lot of research effort has been dedicated to the automated analysis of medical imaging data and to the question of how Computer-Aided Detection and Diagnosis algorithms can assist the tasks mentioned above. However, one of the biggest challenges in this regard is the highly heterogeneous nature of medical imaging data. The acquisition of data from different imaging modalities, like X-ray or Magnetic Resonance Imaging (MRI), changes of acquisition parameters, and the use of different scanners results in diverse data. The varying spatial resolution as well as the high dimensionality of the data pose additional challenges to the development of automated solutions. In this thesis, we investigate different machine learning-based methods to address the analysis of heterogeneous medical imaging data, such as multi-parametric, multi-modal, multi-center, or multi-view data. We present three different pipeline approaches that follow generalization- and fusion- based approaches and demonstrate their applicability on diverse public datasets. Our contributions target two selected use cases in radiology: the semantic labeling of the spine in MRI data and the analysis of mammograms. In semi- and fully-automated spine labeling in MRI data, we are confronted with the problem that MRI data does not exhibit a standardized intensity scale, which results in a large variety of different image contrasts. To overcome this problem for semantic spine labeling, we propose an iterative labeling pipeline that employs Entropy-Optimized Texture Models (ETMs). The application of trained ETMs allows us to apply our models to a wide range of different MRI data. This is in contrast to various related works that develop methods for specific MRI image sequences and protocols. For the analysis of mammography screening data, not only one but four X-ray images from different fields of view are available that form a study of a patient. In addition to this multi-view data, we deal with multi-scale information at various levels, e.g., on a patient, image, or lesion level. To utilize and combine this information efficiently, we develop several deep learning-based models that aim for a specific task important in examining mammograms, such as the localization of abnormalities. For a comprehensive prediction on a patient level, we propose to fuse predictions and features from the individual models to increase performance, which is in contrast to standard ensembling techniques. The results in this thesis demonstrate that considering the different aspects of heterogeneous medical imaging data is inevitable to improve both generalization and predictive capabilities of Computer-Aided Detection and Diagnosis methods.