Comprehensive Summary
This study by Xue et al. investigates the use of a novel deep learning model, DCE-UNet, made for the automatic segmentation of x-rays of adolescents with multiple spinal injuries. The model was tested and trained on a spinal x-ray data set that was split into two subsets– anteroposterior view and lateral view. The model's results were validated against existing methods to assess accuracy. The model was found to have a very strong performance with a Dice score of 91.3%, a mIoU of 84.1, and a Hausdorff distance of 4.0. DCE-UNet outperformed other existing models, excelling in its ability to handle images with multiple injuries as well as more complex spinal structures. When compared to similar models, DCE-UNet showed lower computational costs, without a reduction in accuracy. Although DCE-UNet showed lots of promise, the model is still limited in its scope of training data, 2D-only use, and a lack of more robust external validation. The model was only trained in data from a single source, including all 2D images, which limits its overall applicability to a broader range of images.
Outcomes and Implications
Spinal segmentation models have existed for a couple years; however, these models are limited in their accuracy and robustness, creating a space for DCE-UNet to prevail. The adolescent period of spinal growth is crucial in identifying spinal disorders that may be cemented for life, so early detection and prevention of such injuries is of utmost importance. DCE-UNet has clear clinical applications, but a few limitations still serve as roadblocks for integration into the clinical setting. This model still serves as only a diagnostic tool, rather than having capabilities that include clinical actions and treatment steps. More validation tests with external data must be performed before this model is ready to be considered as an integral clinical tool.