Extraction of time-dependent properties from medical ultrasound image series

Medical ultrasound offers the unique possibility to gather real-time image series, providing insights into dynamic processes of the human body. The interpretation of the acquired sequences, however, can be challenging, especially when the dynamic property of interest is superimposed by, for example, respiratory motion. This thesis investigates how to automate the extraction of time-dependent properties from motion-affected ultrasound image series considering two concrete use cases. The first part deals with extracting two image features from contrast-enhanced ultrasound (CEUS) acquisitions of liver lesions, which are relevant for diagnosis. Both features characterise the distribution of the contrast agent in the lesion compared to normal liver tissue over time. Deep learning-based classifiers are exploited on a large collection of 500 labelled heterogeneous CEUS acquisitions. The influence of aspects such as motion compensation and data representation on the classification result is systematically analysed. In the second part, a use case from physiotherapy is explored in which segmental stabilising exercises are incorporated to treat low back pain. During those exercises, the contraction status of the abdominal muscles can be monitored via ultrasound imaging. Automating the extraction of this status has the potential to enable wearable ultrasound biofeedback devices which can be used for example during home training. Several deep learning-based segmentation algorithms for the three relevant abdominal muscles are evaluated, using time series acquired from volunteers performing exercises.Also, different strategies to assess the contraction state from the obtained segmentations are explored. Both use cases showed that motion can affect the assessment of dynamic image features in ultrasound. Using effective algorithms, it can be controlled to some extent, enabling the use of information along the sequence to retrieve the desired properties.