Feasibility of generating sagittal radiographs from coronal views using GAN-based deep learning framework in adolescent idiopathic scoliosis

Background: Minimizing radiation exposure is crucial in monitoring adolescent idiopathic scoliosis (AIS). Generative adversarial networks (GANs) have emerged as valuable tools being able to generate high-quality synthetic images. This study explores the use of GANs to generate synthetic sagittal radiographs from coronal views in AIS patients. Methods: A dataset of 3,935 AIS patients who underwent spine and pelvis radiographic examinations using the EOS system, which simultaneously acquires coronal and sagittal images, was analyzed. The dataset was divided into training-set (85%, n = 3,356) and test-set (15%, n = 579). GAN model was trained to generate sagittal images from coronal views, with real sagittal views as reference standard. To assess accuracy, 100 subjects from the test-set were randomly selected for manual measurement of lumbar lordosis (LL), sacral slope (SS), pelvic incidence (PI), and sagittal vertical axis (SVA) by two radiologists in both synthetic and real images. Results: Sixty-nine synthetic images were considered assessable. The intraclass correlation coefficient ranged 0.93-0.99 for measurements in real images, and from 0.83 to 0.88 for synthetic images. Correlations between parameters of real and synthetic images were 0.52 (LL), 0.17 (SS), 0.18 (PI), and 0.74 (SVA). Measurement errors showed minimal correlation with scoliosis severity. Mean ± standard deviation absolute errors were 7 ± 7° (LL), 9 ± 7° (SS), 9 ± 8° (PI), and 1.1 ± 0.8 cm (SVA). Conclusion: While the model generates sagittal images visually consistent with reference images, their quality is not sufficient for clinical parameter assessment, except for promising results in SVA. Relevance statement: AI can generate synthetic sagittal radiographs from coronal views to reduce radiation exposure in monitoring adolescent idiopathic scoliosis (AIS). However, while these synthetic images appear visually consistent with real ones, their quality remains insufficient for accurate clinical assessment. Key points: AI can be exploited to generate synthetic sagittal radiographs from coronal views. Dataset of 3,935 subjects was used to train and test AI-model; spinal parameters from synthetic and real images were compared. Synthetic images were visually consistent with real ones, but quality was generally insufficient for accurate clinical assessment.