<p>Scoliosis is a spatial deformity characterized by an abnormal curvature in the spine. The Cobb angle, which is commonly used to quantify the severity of scoliosis, is manually or semi-manually measured by specialists and is normally subjected to inter- and intra-observer variation. Additionally, accurate Cobb angle estimation is vital for clinical diagnosis and assessment. Therefore, a new reliable, fully automated method is needed to support clinical decision-making. The current advancement in deep neural networks provides a great method for automated Cobb angle estimation. DenseNet, a pre-trained deep neural network (DNN), was modified to perform regression by predicting the main thoracic (MT) Cobb angle. The network was trained on two different datasets (automated assessment of scoliosis from the spinal curve estimation Challenge data set and King Abdullah University Hospital dataset), consisting of 797 anterior-posterior (AP) full spinal X-ray images, along with their corresponding Cobb angles, and was later tested to evaluate the model’s performance. The proposed method achieved excellent reliability and robustness. When tested on unseen images from both datasets, the mean absolute error (MAE) was 4.67°, the symmetric mean absolute percentage error (SMAPE) was 12%, and the R-square was 0.89. Additionally, when the trained model was tested on the AASCE2019 only, it performed much better, yielding an MAE of 1.36°, a SMAPE of 4.0%, and an R-squared of 0.98. These results highlight the model’s generalizability and effectiveness in predicting the Cobb angle from the raw X-ray images, with no manual feature extraction needed.We introduced an AI model to predict the Cobb angle from raw X-ray images. The proposed model demonstrated robustness and high accuracy when tested on the two datasets. The resulting deviations from the ground truths were much smaller than those typically associated with manual measurements, indicating its promise for deployment in a clinical setting.</p>