Recurrent multi-view 6DoF pose estimation for marker-less surgical tool tracking

Marker-based tracking of surgical instruments facilitates surgical navigation systems with high precision, but requires time-consuming preparation and is prone to stains or occluded markers. Deep learning promises marker-less tracking based solely on RGB videos to address these challenges. In this paper, object pose estimation is applied to surgical instrument tracking using a novel deep learning architecture.

We combine pose estimation from multiple views with recurrent neural networks to better exploit temporal coherence for improved tracking. We also investigate the performance under conditions where the instrument is obscured. We enhance an existing pose (distribution) estimation pipeline by a spatio-temporal feature extractor that allows for feature incorporation along an entire sequence of frames.

On a synthetic dataset we achieve a mean tip error below 1.0 mm and an angle error below 0.2 $_{}^{\circ }$ using a four-camera setup. On a real dataset with four cameras we achieve an error below 3.0 mm. Under limited instrument visibility our recurrent approach can predict the tip position approximately 3 mm more precisely than the non-recurrent approach.

Our findings on a synthetic dataset of surgical instruments demonstrate that deep-learning-based tracking using multiple cameras simultaneously can be competitive with marker-based systems. Additionally, the temporal information obtained through the architecture's recurrent nature is advantageous when the instrument is occluded. The synthesis of multi-view and recurrence has thus been shown to enhance the reliability and usability of high-precision surgical pose estimation.