Associate Professor Department of Neurosurgery, University of Pennsylvania Philadelphia, Pennsylvania, United States
Introduction: Biomedical imaging provides a wealth of information about individual patients and populations. Drawn from imaging, standardized metrics can improve population management and support care prediction modeling. However, the volume of data available can make analysis challenging. Automatic segmentation algorithms offer new tools for image analysis. Neural network (NN) algorithms have segmented individual vertebral bodies in MRI and CT images, but few studies have demonstrated whole spine segmentation in vivo. Our objective is to use Mask R-CNN, a convolutional NN, to make whole spine segmentation possible.
Methods: Whole spine imaging scans (EOS) from 59 adult patients (128 sagittal and coronal planes) with mild to severe scoliosis were obtained. Human vertebral columns were manually segmented to serve as the ground truth. Manual segmentations were then split into training and validation sets in a 4:1 ratio. Sagittal and coronal segmentation pre-trained Mask R-CNN models were trained first in a training set and subsequently tested on a validation set. The exact accuracy of the AI-generated segmentations to the ground truth segmentations was determined using the DICE score, an index of similarity (0-1.0).
Results: Qualitatively, the structure of the vertebral column was well-segmented by both models. On validation images, the trained models achieved moderate similarity to the ground truth (average DICE similarity scores of 0.64 and 0.68 for sagittal and coronal planes respectively). Further, the strongest DICE similarity score for sagittal and coronal segmentations were 0.81 and 0.88, respectively.
Conclusion : Our Mask R-CNN model successfully segmented the vertebral column within sagittal- and coronal-plane EOS scans. In the future, we plan to test our state-of-the-art computer vision algorithms across a larger training dataset. Once completed, we hope to automate whole-spine segmentation to increase diagnostic efficiency and decrease costs.
