Markerless 6D Pose Estimation of Surgical Tools and Anatomical Shapes from RGB-D images: a comparison of two approaches based on synthetic data

5 pages•Published: December 17, 2024

Basile Longo, Salaheddine Sta, Éric Stindel and Guillaume Dardenne

Abstract

In the realm of Computer Assisted Surgery (CAS), the accurate localization of anatom- ical features and surgical tools is crucial for enhancing surgical outcomes. Important efforts have been recently focused toward markerless navigation to reduce intraoperative intrusiv- ity. Unfortunately, recent studies don’t always satisfy the required accuracy and precision to be used clinically. This research investigates the application of deep learning models for pose estimation of synthetic anatomical features and surgery instruments to improve the localization accuracy. Based on the models from CosyPose and Coupled-Iterative- Refinement methods, we applied a three-step pose estimation process. Using 3D meshes and BlenderProc we generated a synthetic RGB-D dataset of scenes including tibias, fe- murs, shoulder’s glenoids and surgical instruments (Cutting guide) to train and test our model. Moreover, we compared our results to the Point Pair Features (PPF) method, a conventional pose estimation algorithm based on point cloud data. We found a signifi- cant enhancement regarding the accuracy with our model, achieving sub-millimetric and sub-degree accuracy, surpassing the PPF algorithm. We also provided qualitative exam- ples of the estimated poses to visualize the accuracy of our model. While the proposed method shows promising results, challenges remain in particular passing from synthetic to real-world data. Future efforts will focus on collecting annotated real-world data.

Keyphrases: deep learning, pose estimation, rgb d sensors, surgical navigation

In: Joshua W Giles and Aziliz Guezou-Philippe (editors). Proceedings of The 24th Annual Meeting of the International Society for Computer Assisted Orthopaedic Surgery, vol 7, pages 145-149.

Links:	https://easychair.org/publications/paper/VGZX7
	https://doi.org/10.29007/z3td

BibTeX entry

@inproceedings{CAOS2024:Markerless_6D_Pose_Estimation,
  author    = {Basile Longo and Salaheddine Sta and Éric Stindel and Guillaume Dardenne},
  title     = {Markerless 6D Pose Estimation of Surgical Tools and Anatomical Shapes from RGB-D images: a comparison of two approaches based on synthetic data},
  booktitle = {Proceedings of The 24th Annual Meeting of the International Society for Computer Assisted Orthopaedic Surgery},
  editor    = {Joshua W Giles and Aziliz Guezou-Philippe},
  series    = {EPiC Series in Health Sciences},
  volume    = {7},
  publisher = {EasyChair},
  bibsource = {EasyChair, https://easychair.org},
  issn      = {2398-5305},
  url       = {/publications/paper/VGZX7},
  doi       = {10.29007/z3td},
  pages     = {145-149},
  year      = {2024}}

Download PDF Open PDF in browser