Intraoperative 3D reconstruction from sparse arbitrarily posed real X-rays

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Navigation

Background:

• Spine surgery requires precise navigation, often using image-based systems.
• Supervised learning for 3D anatomy reconstruction from fluoroscopy has limitations, including data requirements and generalization issues.
• Instance-learning methods like Gaussian splatting offer potential but haven't been applied to sparse intraoperative X-rays.

Purpose of the Study:

• To extend Gaussian splatting for reconstructing anatomically consistent 3D volumes from sparse, arbitrarily posed intraoperative X-rays.
• To improve reconstruction quality and visual consistency using an anatomy-guided radiographic standardization technique.
• To demonstrate a pretraining-free, adaptable framework for intraoperative 3D imaging.

Main Methods:

• Extended the [Formula: see text]-Gaussian splatting framework for 3D volume reconstruction.
• Implemented an anatomy-guided radiographic standardization using style transfer for improved visual consistency.
• Evaluated the approach on an ex-vivo dataset, assessing clinical utility and quantitative metrics.

Main Results:

• Achieved anatomically consistent 3D reconstructions from sparse-view X-rays without pretraining.
• Standardization step enhanced anatomical clarity and reconstruction quality.
• Expert surgical evaluation confirmed clinical utility for navigation with 20-30 views.
• Quantitative metrics showed performance trade-offs but validated standardization benefits.

Conclusions:

• Demonstrated the feasibility of instance-based 3D volumetric reconstruction from arbitrary sparse-view X-rays.
• The proposed framework advances intraoperative 3D imaging for spine surgery navigation.
• The method is adaptable to new patients and anatomies without pretraining.