Localizing Neurosurgical Instruments Across Domains and in the Wild
Proceedings of the Fourth Conference on Medical Imaging with Deep Learning, PMLR 143:581-595, 2021.
Towards computer-assisted neurosurgery, robust methods for instrument localization on neurosurgical microscope video data are needed. Specifically for neurosurgical data, challenges arise from visual conditions such as strong blur and from an unknowingly large variety of instrument types. For neurosurgical domain, instrument localization methods must generalize across different sub-disciplines such as cranial tumor and aneurysm surgeries which exhibit different visual properties. We present and evaluate a methodology towards robust instrument tip localization for neurosurgical microscope data, formulated as coarse saliency prediction. For our analysis, we build a comprehensive dataset comprising in-the-wild data from several neurosurgical sub-disciplines as well as phantom surgeries. Comparing single stream networks using either image or optical flow information, we find complementary performance of both networks. Plain optical flow enables better cross-domain generalization, while the image-based network performs better on surgeries from the training domain. Based on these findings, we present a two-stream architecture that fuses image and optical flow information to utilize the complementary performance of both. Being trained on tumor surgeries, our architecture outperforms both single stream networks and shows improved robustness on data from different neurosurgical sub-disciplines. From our findings, future work must focus more on how to incorporate optical flow information into fusion architectures to further improve cross-domain generalization.