AI Image Analysis for Hounsfield Units in Preoperative Spine CT scans: An Automated Screening for Osteoporosis in Patients undergoing Spine Surgery

Introduction: Despite being associated with poor surgical outcomes, osteoporosis is commonly investigated and treated prior to an elective spine surgery. Preoperative optimization of bone metabolism can potentially enhance bone healing and therefore improve surgical outcomes. We aimed to develop an AI model that can detect Hounsfield units (HU) at L1 vertebra from a spine CT scan. The objective is to enable identification of osteoporosis cases through imaging, facilitating timely management of the disease.

Methods: A database comprising 523 thoracolumbar CT scans was used to develop the model. The data set included normal CT scan images from diverse patient demographics. A model for segmentation of the vertebrae was developed using a 70:15:15 division of the total cohort. The model was then dictated to mark square in the sagittal image of L1 vertebra with minimum 1 mm distance from the outer vertebral boundary to exclude the cortical bone. The subsequent square area was used to determine the average HU units of the patient. The model was evaluated using standard metrics, including Dice scores, to assess its performance.

Results: The model was trained on 367 cases, followed testing and validation with 78 cases each. A 3D full resolution UNet model was trained utilizing the nnUNet framework to conduct comprehensive volumetric segmentation on the amalgamated dataset. This approach aimed to harness the diversity and specificity of the combined datasets to enhance the robustness and accuracy of the segmentation model. The developed model demonstrated exceptional performance, achieving a Dice score of 0.95 for the L1 vertebra, signifying a high level of spatial overlap between predicted and ground truth regions. Additionally, for other vertebrae, the model exhibited a Dice score of 0.80. These results highlight the model's robustness and effectiveness in automated detecting of Hounsfield units.

Conclusion : Our model can accurately predict the Hounsfield units of the patients, allowing for automation of an important, and commonly overlooked, evaluation of preoperative bone quality. Consequently, patients with high risk of osteoporotic complications can be referred for preoperative bone quality optimization.