Automatic laminectomy cutting plane planning based on artificial intelligence in robot assisted laminectomy surgery

Zhuofu Li, Yonghong Zhang, Chengxia Wang, Shanshan Liu, Xiongkang Song, Xuquan Ji, Shuai Jiang, Woquan Zhong, Lei Hu, Weishi Li

TL;DR

Researchers have developed SPU-Net, a deep learning framework that automates the planning of cutting planes for robot-assisted laminectomy. By accurately identifying seven anatomical landmarks on CT scans, the system generates personalized surgical paths with sub-millimeter precision, potentially reducing the steep learning curve and risks associated with spinal decompression surgery.

What Problem Does This Paper Address?

Lumbar spinal stenosis (LSS)—a narrowing of the spinal canal—often requires a laminectomy, where a portion of the vertebral bone (the lamina) is removed to relieve pressure on nerves. This is a high-stakes "game of millimeters." Cut too little, and the decompression is inadequate; cut too much or in the wrong place, and you risk iatrogenic instability or catastrophic damage to the dural sac and nerve roots.

While surgical robots currently excel at assisting with pedicle screw placement, they lack the "intelligence" to plan laminectomy cuts autonomously. Currently, surgeons must manually define these paths based on years of experience. This paper aims to bridge that gap by using AI to automate the planning process, making the procedure safer and more standardized.

Methods

The authors propose a two-stage "Automatic Laminectomy Cutting Plane Planning" (ALCPP) workflow:

  1. Landmark Identification (SPU-Net): The team developed the Spatial Pyramid Upsampling Network (SPU-Net). Unlike standard U-Nets that use simple pooling and upsampling (which can lose spatial information), SPU-Net incorporates 3D Patch Merging and 3D Patch Expanding layers borrowed from transformer architectures. These layers are learnable, allowing the network to downsample and upsample while preserving critical features. A "Spatial Pyramid" structure then fuses multi-scale features to output heatmaps for seven specific vertebral landmarks.
  2. Coordinate System Fitting: Once the landmarks (like the edges of the pedicles and the vertebral body) are located, the system fits a personalized 3D coordinate system to the specific vertebra.
  3. Plane Generation: Using clinical "rules of thumb" (e.g., placing longitudinal cuts at 75% of the distance between the midline and the pedicle edge), the algorithm automatically generates three precise cutting planes.

Datasets

The study utilized the VerSe2020 dataset, a large-scale, open-source benchmark for spinal segmentation and labeling.

  • Training/Validation: 217 CT scans were processed into 1,042 individual vertebral images.
  • Testing: 64 CT scans (260 images for landmark testing; 320 vertebrae for plane planning evaluation).
  • Ground Truth: Landmarks were manually annotated by an attending physician with two years of experience following a standardized operating procedure (SOP).

Key Insights and Ideas

  • Landmarks over Segmentation: Most previous attempts at this problem tried to segment the entire lamina. However, the lamina is highly irregular and varies wildly between patients. By focusing on seven discrete landmarks instead, the authors created a more robust system that works even when the bone structure is "noisy" or deformed (e.g., in cases of spina bifida).
  • Information Preservation: The move away from traditional max-pooling to "Patch Merging" is a significant technical choice. It treats the downsampling process as a learnable operation, ensuring that the fine-grained spatial details required for sub-millimeter surgical planning aren't discarded.
  • Clinical Logic Integration: The AI doesn't just "guess" where to cut; it uses the landmarks to apply established surgical wisdom algorithmically, ensuring the output is immediately interpretable and useful to a surgeon.

Key Results

The system demonstrated high accuracy and clinical viability:

  • Localization Accuracy: SPU-Net achieved an average landmark localization error of 0.65 ± 0.70 mm, which is well within the safety margins required for spinal surgery.
  • Planning Quality: The researchers used a grading system (A, B, C) to evaluate the generated planes.
  • Longitudinal Planes: 97.18% were rated as "Grade A" (Excellent).
  • Transverse Planes: 99.38% were rated as "Grade A."
  • Robustness: The method remained effective in cases of skewed spinous processes and anatomical variations like spina bifida, where traditional segmentation-based methods often fail.

Limitations and Future Work

Despite the strong results, the authors highlight a few hurdles before this hits the OR:

  • Data Diversity: The dataset primarily consisted of "normal" spinal CTs. The algorithm's performance on severe spinal deformities (scoliosis, etc.) remains a question mark.
  • Noise Sensitivity: High levels of CT image noise can still lead to landmark deviation and planning failure.
  • Human-in-the-loop: The system is currently in the laboratory stage. The authors emphasize that any future clinical version must retain an interface for surgeons to inspect and manually correct the AI's plan.

Future research will focus on expanding the dataset with more pathological cases and implementing automated noise-reduction techniques to improve the system's robustness.

Summary updated

Abstract

Objective: This study aims to use artificial intelligence to realize the automatic planning of laminectomy, and verify the method. Methods: We propose a two-stage approach for automatic laminectomy cutting plane planning. The first stage was the identification of key points. 7 key points were manually marked on each CT image. The Spatial Pyramid Upsampling Network (SPU-Net) algorithm developed by us was used to accurately locate the 7 key points. In the second stage, based on the identification of key points, a personalized coordinate system was generated for each vertebra. Finally, the transverse and longitudinal cutting planes of laminectomy were generated under the coordinate system. The overall effect of planning was evaluated. Results: In the first stage, the average localization error of the SPU-Net algorithm for the seven key points was 0.65mm. In the second stage, a total of 320 transverse cutting planes and 640 longitudinal cutting planes were planned by the algorithm. Among them, the number of horizontal plane planning effects of grade A, B, and C were 318(99.38%), 1(0.31%), and 1(0.31%), respectively. The longitudinal planning effects of grade A, B, and C were 622(97.18%), 1(0.16%), and 17(2.66%), respectively. Conclusions: In this study, we propose a method for automatic surgical path planning of laminectomy based on the localization of key points in CT images. The results showed that the method achieved satisfactory results. More studies are needed to confirm the reliability of this approach in the future.
BibTeX
@misc{li2023automatic,
  author = {Zhuofu Li and Yonghong Zhang and Chengxia Wang and Shanshan Liu and Xiongkang Song and Xuquan Ji and Shuai Jiang and Woquan Zhong and Lei Hu and Weishi Li},
  title = {{Automatic laminectomy cutting plane planning based on artificial intelligence in robot assisted laminectomy surgery}},
  year = {2023},
  eprint = {2312.17266},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.17266},
}
Figure