Introduction: Low-dose radiographic techniques are increasingly implemented to reduce patient exposure to ionizing radiation; however, even minimal doses entail residual risk, warranting continued efforts toward dose minimization. Recent advances in artificial intelligence, particularly deep learning, have introduced potential strategies to address this challenge. Prior investigations employing generative adversarial networks to reconstruct sagittal radiographs from coronal projections for the evaluation of sagittal spinal parameters yielded suboptimal accuracy, limiting clinical applicability. The present study aimed to assess an alternative deep learning framework designed to directly estimate sagittal parameters from coronal radiographs.
Methods: A retrospective cohort of 449 adults without spinal deformities who underwent biplanar spine and pelvic imaging with the EOS system was analyzed. This system enables simultaneous acquisition of coronal and sagittal views. Six sagittal parameters—thoracic kyphosis (TK), lumbar lordosis (LL), sacral slope (SS), pelvic incidence (PI), pelvic tilt (PT), and sagittal vertical axis (SVA)—were manually measured on sagittal images. The dataset was partitioned into training (70%), validation (15%), and test (15%) subsets. A multi-output regression neural network was trained to predict these parameters from coronal images. Model architectures and hyperparameters were optimized using validation data. Performance was evaluated on the test set using mean absolute error and correlation coefficients.
Results and Discussion: Mean absolute errors ranged from 6±6° to 9±6° for TK, LL, SS, and PI, 5±4° for PT, and 2.2±2.0 cm for SVA. Maximum deviations reached 26°–40° for TK, LL, SS, and PI, 23° for PT, and 8.8 cm for SVA. Correlation coefficients were 0.71 (TK), 0.64 (LL), 0.43 (SS), 0.28 (PI), 0.73 (PT), and 0.76 (SVA).
Conclusions: Despite moderate correlations for selected parameters, predictive accuracy was insufficient for clinical translation. These findings reinforce prior evidence, indicating that sagittal alignment cannot be reliably inferred from coronal radiographs, likely due to the inherent lack of sagittal-plane information.