Abstract:

This paper is the extension of a detailed anatomical model (Sachse et al., 1996a) (Sachse et al., 1996b) with the three-dimensional orientation of skeletal muscle fibres (Figure 1). The orientation is interpolated basing on two sets with restrictions of different types. The first set consists of points for which the orientation is known. The second set consists of points with an assigned normal of orientation. These sets are created by detection with manual or automatic methods using techniques of digital image processing. The interpolation works iteratively employing the averaging orientations in the 6-neighbourhood. The average of neighbouring orientations is calculated by determination of their principal axis.

Abstract:

Background: In neurosurgeries today it is common to use surgical microscopes in order to enable surgical interventions on millimetre scale. A defocused view in the region of interest can restrict the ability of the surgeon to act. Because the instrument tips were found to be regions of interest, this thesis explores blur detection at instrument tips for neurosurgical images. To account for sub-domain shifts (e.g. different hospitals, surgeries, instruments) in neurosurgery, an emphasis is put on performance evaluations across neurosurgical sub- domains. Methods: As of now, no medical blur data sets are available. Therefore, synthetic and real blur data sets are generated using a simulation environment and a surgical microscope in a laboratory set-up. Existing non-medical blur data sets suffer from perception-bias introduced during human annotation. This thesis presents two data generation methods for the acquisition of physically-based blur annotations at instrument tips. Generated image data is split into three sub-domains: one big data set containing synthetic images and two smaller data sets consisting of real images. All three data sets represent different visual properties to enable cross-domain performance evaluations. A context-specific approach, using Convolutional Neural Networks, was developed and the Laplacian Variance was chosen as a context-independent approach for a comparative analysis. Results: Trained on synthetic data, the context-specific approach shows better in-domain and out-of-domain performance than the context-independent approach. Additionally, the context-specific approach, trained on synthetic data only, shows better performance on unseen real data sets, even if the context-independent approach used these real data sets for self-optimization. In this case, the out-of-domain performance of the context-specific approach is higher than the in-domain performance of the context-independent approach. Conclusion: The context-specific approach demonstrated workability and substantial perfor- mance across neurosurgical sub-domains. While it was observed that the context-independent approach uses features irrelevant to the task, the context-specific approach can benefit from its ability to build internal representations and context understanding. It is assumed that this results in better generalisation across sub-domains, even if those sub-domains remain unknown.