TL;DR
This paper introduces a novel markerless optical tracking system for navigated spine surgery. By directly detecting and matching anatomical spine features in real-time, it aims to overcome the limitations of current marker-based systems, achieving sub-millimeter accuracy and simplifying the surgical workflow.
What Problem Does This Paper Address?
Current surgical navigation systems for spinal procedures heavily rely on Dynamic Reference Frames (DRFs) or skin markers to track patient movement. However, these methods have drawbacks: DRFs attached to a single vertebra can be inaccurate due to intervertebral movement, and markers can become dislodged or obscured. This can lead to incorrect navigation information, increasing the risk of neurovascular injuries and reoperations. The paper addresses the need for a more robust and accurate patient tracking solution that directly monitors the spine itself, without relying on external markers.
Methods
The core of this research is a computer vision framework designed to detect and track anatomical spine features. The process involves:
- Image Preprocessing: This includes image rectification to align epipolar projections from stereo camera views and contrast enhancement using CLAHE (Contrast Limited Adaptive Histogram Equalization) to improve feature visibility. Spine segmentation is then performed using Otsu's method, followed by manual cropping to focus on the spinal region.
- Feature Detection and Matching: Four established feature detection algorithms were evaluated:
- SURF (Speeded Up Robust Features): Detects interest points based on Hessian matrix determinants and uses Haar wavelets for orientation.
- MSER (Maximal Stable Extremal Regions): Identifies stable regions across varying intensity thresholds, offering affine invariance.
- FAST (Features from Accelerated Segment Test): A fast corner detection algorithm that compares pixel brightness against a threshold.
-
ORB (Oriented FAST and Rotated BRIEF): Combines FAST for detection with BRIEF for descriptor extraction, adding rotation invariance. These algorithms identify salient features on the spine. Normalized cross-correlation is used for feature matching between stereo views.
-
3D Triangulation: Once corresponding 2D features are found in multiple views, 3D triangulation is employed to determine their precise spatial location. The method minimizes 3D error by selecting the midpoint between back-projected image points.
Datasets
The study utilized optical images acquired from 23 retrospective open spinal surgery cases. These images were captured by two gray-scale cameras as part of an existing Augmented Reality Surgical Navigation (ARSN) system. No new benchmark datasets were introduced; instead, the framework was validated on real-world clinical data.
Key Insights and Ideas
- Markerless Tracking is Feasible: The study demonstrates that directly tracking anatomical spine features is a viable alternative to marker-based systems, offering comparable or superior accuracy.
- Feature Detection Algorithms Matter: The choice of feature detection algorithm significantly impacts both the number of detected features and the resulting 3D triangulation error.
- Trade-offs Exist: While ORB detected the highest number of features, SURF and FAST provided lower triangulation errors. This suggests a potential for hybrid approaches.
- Illumination is Critical: Poor lighting conditions were identified as a significant challenge, impacting feature detection and potentially leading to inaccurate matching.
- Sub-millimeter Accuracy Achieved: The framework consistently achieved a 3D triangulation error of less than 0.5 mm, which is clinically acceptable for navigated spine surgery.
Key Results
- Accuracy: The 3D triangulation error for matched features was consistently below 0.5 mm across all tested algorithms. SURF and MSER slightly outperformed others in terms of mean triangulation error.
- Feature Count: ORB detected the highest number of matched features (average of 131 inliers), while SURF and FAST provided a more consistent number of inliers (average of 40 and 28, respectively), which is crucial for reliable transformation estimation.
- Speed: FAST was the fastest feature detection algorithm, achieving 33 frames per second (fps), making it suitable for real-time applications. SURF also performed well at 20 fps. MSER and ORB were significantly slower.
- Validation: The framework was validated on 23 patients, demonstrating its robustness in a clinical setting.
Limitations and Future Work
- Real-time Operation: The current system operates on pairs of frames, not in true real-time. Future work will focus on optimizing hardware and software for real-time spine feature localization.
- Illumination Control: The study acknowledges that poor illumination can affect performance. Future applications will require controlled lighting conditions in the operating room.
- Visibility Challenges: In minimally invasive surgery, spine visibility is limited. While the current approach is suitable for open surgery, advancements in camera technology (e.g., color or hyperspectral) might be needed for broader applicability.
- Intervertebral Movement: The paper suggests future work to reconstruct and evaluate intervertebral movement for more accurate intraoperative navigation, potentially by fusing data with intraoperative CT scans.
- Algorithm Combination: The authors propose combining SURF and ORB features to potentially achieve higher matching accuracy and robustness.
Summary updated