Hip osteoarthritis (HipOA) is an increasingly prevalent disease in today’s hyper-aged society. Its diagnosis is usually based on radiographs and requires the clinical expertise to grade the hip deformity and disease progression based on Kellgren and Lawrence (KL) and Crowe gradings. However, the diagnosis is subjective and depends on the surgeon, which may introduce inter– and intra– observer variabilities. Widely used deep learning models, including a recently proposed model (VisionTransformer (ViT)), were attempted in classification and regression settings. For validation, a database of 394 unilateral digitally- reconstructed radiographs (DRRs), generated from CT images of the hip region of 197 HipOA patients was used. The KL and Crowe grades were combined into a single label encoding one of seven combinations of the two, thus representing the disease severity. The grading accuracy was assessed using exact class accuracy (ECA) and one-neighbour class accuracy (ONCA). The largest ECA was obtained for ViT model with 0.656±0.015 for regression. By tolerating one-neighbor failure, ViT accuracy could increase to 0.962±0.011 for regression. Additionally, ViT produced the smallest mean error between the true and predicted labels in the regression (continuous label), which was 0.417 (IQR: 0.740). In general, DRRs of the normal to mild stages had higher accuracy compared with those of severe stages. Future work will focus on increasing training data of severe stages and analyze disease progression in large-scale databases.