Introduction: Analysing intraoperative visual findings (such as vessels, nerves, the convoluted borders of tumor cavities, and other clinically significant or interesting features) in a three-dimensional context, along with spatially correlating these findings to postoperative CT and MRI scans, has long posed challenges for neurosurgeons. Current techniques often fall short in reconstructing intraoperative images or videos for superimposition onto postoperative scans. Gaussian splatting, a recent advancement in neural rendering, allows for effective 3D reconstruction from standard video or image inputs. This technique may enhance postoperative analysis by facilitating the superimposition of intraoperative visuals onto postoperative images, thereby improving visuo-spatial correlation between intraoperative findings and postoperative scans. This study evaluates the application of Gaussian splatting to create a 3D model of the operative findings, which is subsequently overlaid on postoperative scans for a more comprehensive assessment.
Methods: Intraoperative video footage was systematically captured during clinically significant steps of the procedure or at the end of surgery. Additionally, images were acquired as needed. After preprocessing the media, a point cloud was generated using a Gaussian splatting algorithm implemented in a Python-based application. The 3D reconstructed model was then created and further processed in Blender 4.0. Simultaneously, 3D models of the postoperative scans were generated using 3D Slicer 5.0.3 software. Both models were superimposed and aligned to produce a final composite model for analysis.
Results: The Gaussian splatting technique successfully reconstructed a detailed 3D model of the tumor cavity, preserving critical spatial characteristics. When overlaid on postoperative scan, the superimposed model provided enhanced visualization of resection margins of tumor, fracture lines in TBI, vessels, etc. providing ultra-realistic 3D models with visually accurate spatial co-relation. In 95% of cases, the boundaries identified in the splatting reconstruction matched those seen on MRI within a 2 mm margin of error. This underscores the clinical utility of superimposed models in postoperative period.
Conclusion : This study demonstrates that Gaussian splatting can be effectively used as augmented visualization tool with accurate superimposition on postoperative scans.
.jpg)
