Segmentation peropératoire des vertèbres à partir de nuages de points 3D pour le suivi sans rayonnement de la courbure vertébrale dans la chirurgie de la scoliose

Posterior spinal fusion surgery (PSFS) typically involves two stages: placement of pedicle screws to anchor the vertebrae, followed by spinal realignment to gradually correct the scoliosis deformity. While navigation systems have improved the accuracy of screw placement, the realignment phase remains largely dependent on the surgeon’s judgment and experience. Intraoperative radiographs are used sparingly, primarily at the end of surgery to evaluate the final realignment, limiting their role in continuous feedback and raising concerns about radiation exposure for patients and staff.

Our objective is to enable continuous intraoperative assessment of vertebral alignment in PSFS without radiation. RGB-D structured-light sensing, yielding direct 3D point clouds, can be used during surgery to capture the exposed anatomy. Such sensors are already used in the operating room, for example, an RGB-D sensor is integrated into the 7D Surgical Navigation Platform (SeaSpine, an Orthofix Company, Lewisville, TX, USA). However, these point clouds are often large, noisy, and include skin tissue, draping, and tools, making it challenging to isolate vertebral structures. In this work, we developed a deep learning-based framework that segments vertebrae from intraoperative point clouds in real time. Training Point Transformer V3 with stratified downsampling and color-robust augmentation achieved a 14.2% improvement in segmentation accuracy over prior RGB-D-based methods on the public SpineDepth dataset; a Wilcoxon signed-rank test confirmed this gain was statistically significant (p=0.0078). Moreover, extending training with combinations of real and semi-synthetic datasets yielded satisfactory qualitative performance on real PSFS acquisitions.

These findings highlight the potential of integrating deep learning into scoliosis surgery workflows. By segmenting vertebrae intraoperatively, preoperative 3D spine reconstructions can be registered to real-time point cloud data, enabling frequent, radiation-free tracking of spinal alignment throughout surgery. This capability could provide surgeons with continuous feedback during realignment, reduce reliance on radiographs, and ultimately enhance patient safety and outcomes.