Self-Supervised Learning for Improving Automated Predictions of Cobb Angles from X-Ray Images: Machine Learning Study (Preprint)

Abstract

Scoliosis is an abnormal spine curvature with a frontal plane deviation of more than 10 degrees and axial rotation. The measuring of the Cobb angle is time-consuming. It contributes to observer disparities due to variations in identifying the end vertebra, posing a challenge for inexperienced clinicians in accurately assessing scoliosis. SSL tasks expand the dataset before pretraining, improving the model’s efficiency at learning unlabeled features of data. To explore the use of SSL to address this issue, we used a dataset which comprises 609 x-ray spinal images depicting various scoliosis cases, each annotated with three Cobb angles (MA, TA, and BA). We applied SSL in the form of an image rotation pretext task to train a MobileNetV2 model to understand aspects of the images without any training data. We then fine-tuned this model against the same network without any SSL pre-training. We found that the model pre-trained with SSL outperformed the model without SSL pre-training, whether or not the model was initialized using ImageNet weights. This work demonstrates the promise of SSL for Cobb angle prediction using x-ray images. We recommend that the medical imaging research community explore further SSL pre-training strategies for regression tasks such as Cobb angle prediction.