Comprehensive Summary
A method for 3D volumetric segmentation, labelling, and registration of spinal vertebrae was developed and evaluated utilizing a 3D U-Net neural network approach. Using residual and dense interconnections, and a series of network components like activation functions, optimisers, and pooling operations, an algorithm was constructed and trained on raw CT scans from a VerSe’20 dataset. The result was an anatomically aligned 3-D vertebra model. By comparing the 3D model to the true anatomy of the spinal positions, a Dice coefficient of ~92.4% was achieved along with a Hausdorff distance of 5.26 mm. This outperformed every previous model tested on the VerSe’20 dataset, and the full algorithm pipeline was completed in 90–210 seconds. It is important to note that the VerSe’20 dataset contained few CT scans with significant spinal deformity. This could impact the algorithm's ability to model such patients.
Outcomes and Implications
Quickly and accurately identifying the boundaries of spinal vertebrae is essential to efficient pre-operative planning. Current methods of manual analysis, however, are time-consuming and expensive. By implementing the article's 3D U-Net neural network approach, conducting spinal surgeries would become quicker, less resource-intensive, and far more efficient. According to the author, the segmentation-and-registration system has already been implemented into a VR visualization software and is currently being used for pre-operative planning in the clinical setting. Future studies would further validate the algorithm to allow direct intra-operative use.