Trained nnU-Net model for semantic segmentation of human adult cervical vertebrae from CT-Scans

Automatic segmentation of the mid to low cervical spine often shows poor performance, which is detrimental to the development of patient-specific models for numerical simulations. We hypothesised that training a semantic segmentation model specifically on the cervical spine, rather than the full spine as it is usually done, would lead to better results. We trained and validated two models (nnU-Net v.2 convolutional neural network) on 172 computed tomography (CT) images: one segmenting only the cervical spine, and one segmenting the full spine. These models were then tested on an independent set of 268 CT images, unrelated to those selected for model training and internal validation. The DICE metric of the cervical model was 0.951 ± 0.051 and its 95^th percentile Hausdorff's distance (H95) was 1.43 ± 1.44 mm on the test dataset. Both models presented similar performance results (p > 0.05), except for the H95 metric on the test set where the cervical model performed better (p = 0.037). Both models performed better and more homogeneously across vertebral levels than those of the literature. These results might be attributed to a better balance in the number of vertebrae per vertebral levels used for training in both models, rather than a specialisation in segmenting only a specific spine segment.Clinical Relevance- Results further highlight the importance of class balancing in semantic segmentation. The proposed model can be used to develop patient-specific models for numerical simulations, useful for both the fundamental studies of spine biomechanics and for surgery planning. Semantic segmentation of the cervical spine could also eventually assist with medical images interpretation.