Kurzfassung | To tackle the enormous physical and cognitive demands of neurosurgery, several approaches focus on the vision of computer-assisted surgery. Most of these approaches, such as predicting the instrument tip location, rely on motion features derived from the optical flow. However, although they consider state-of-the-art CNNs when computing the optical flow, they typically utilize pre-trained, domain-unaware implementations. This unawareness causes the OF-CNN to experience a severe drop in performance if applied to neurosurgical data, consequently influencing the performance of the succeeding task. We address this problem by adapting the OF-CNN to the neurosurgical domain. For this, we first analyze the optical flow predictions of a pre-trained OF-CNN. By focusing on the impact of prominent visual features on the quality of the optical flow, we derive three features, defocus-blur, reflections of tissue and background motion, which are potentially relevant for adapting the OF-CNN to the neurosurgical domain. Then we design specifically tailored synthetic training datasets including these visual features, which allows us to adapt the pre-trained OF-CNN to the neurosurgical domain. Our evaluation showed that all of our identified visual features are relevant for domain adaptation. Moreover, a suitable combination of features leads to significantly improved accuracy of the optical flow itself and improved usefulness for the high-level vision task of instrument tip localization.
|