PURPOSE: A healthy posture is characterized by symmetric features like equal shoulder height, a straight spinous process line and equally-sized waist triangles. Deviations suggest spinal disorders, among them adolescent idiopathic scoliosis (AIS), a three-dimensional spinal curvature requiring timely treatment. AIS screenings are debated among the scientific community. This study proposes two symmetry-sensitive convolutional neural networks (CNNs) to identify asymmetries in videorasterstereographic back curvature images to classify AIS.

METHODS: A total of 1444 videorasterstereographic measurements were collected from individuals (11 ± 2.4 years), among them 355 with AIS, 306 with other spinal deformities and 783 with healthy posture. Two types of curvature images, Mean and Gauss, were compared using a VGG16 network and a two-channel CNN based on DeepSymNet. The latter analyzes left and right side of the torso image separately, then merges the results and identifies asymmetries.

RESULTS: Both models achieved accuracies between 0.768 and 0.801 during validation, the accuracy of the test set of the tuned DeepSymNet (five repeated trials) was 0.744 ± 0.014, specificity was 0.748 ± 0.025, sensitivity 0.726 ± 0.044 and PPV 0.417 ± 0.013. The equal performances show models were able to learn symmetry, making both suitable for training videorasterstereographic images. The main benefit and greatest challenge was the dataset diversity, which incorporated a variety of postural conditions, confounding AIS features.

CONCLUSION: Adapting CNNs to include symmetry analysis could be an improvement to other known applications of deep learning in spinal deformity analysis. Refining the dataset to incorporate more mild cases could enhance the performance of DeepSymNet in particular. With these modifications, DeepSymNet as well as VGG16 are promising approaches for future AIS screenings.