Exploring Artificial Intelligence Deep Recognition-based Algorithms for Spine 3D Segmentation using Computed Tomography Images: The Next Generation of Spine Pathology Modeling

Introduction: Accurate segmentation processing and subsequent spine 3D modeling is a time-consuming process. While the advancements in deep-learning models and machine-learning algorithms have enabled the automation of some modeling processes, the current impact of the highly adaptable machine-learning algorithms in spine pathology segmentation remains unexplored. This study aims to evaluate the current capacity of artificial intelligence algorithms for spine pathology segmentation based on patient-specific computed tomography (CT) scans.

Methods: The authors combined a current segmentation program (3DSlicer) and the Total Segmentator AI (Artificial Intelligence) plugin model, constructed using the current nnU-Net biomedical deep-based segmentation network. The combination of these automated algorithms was used to assess the segmentation in three spine pathology cases: scoliosis, intradural spinal tumor, and thoracic burst fracture. The metric evaluation included the following scores: IoU, dice, precision, and recall. Epoch modification was used for modeling training. Model analyses were evaluated using AI end-product metric scores and expert opinion comparisons.

Results: Patient-specific anatomical models of various spine pathologies were successfully automated, segmented, and imported into the modeling applications. Anatomical landmarks were accurate and helped spatial anatomical recognition for each 3D model. Findings were supported by experts' opinions while analyzing and comparing the models. Image normalization, data augmentation, residual blocks, and dropout regularization was observed over 80% in our score metrics for the scoliosis, intradural spinal tumor, and thoracic burst fracture models.

Conclusion : This study demonstrates the feasibility of combining AI algorithms into 3D modeling programs to model spinal pathology based on the segmentation of advanced patient imaging. These initial algorithm prototypes may be used for multiple scenarios, including spine surgical planning, ultra-rapid spine modeling, and educational purposes. Future adjustments to the current AI tools will endeavor new avenues to spine research.