Convolutional neural network landmark regression for automated Cobb angle estimation from spinal radiographs

To develop and evaluate a convolutional neural network (CNN)-based coordinate regression model for automated Cobb angle estimation from spinal radiographs. This retrospective study included 460 anteroposterior whole-spine radiographs from patients aged 10-25 years with scoliosis. All radiographs were anonymized and manually annotated with 68 vertebral landmark coordinates. A CNN consisting of four convolution-pooling blocks and four fully connected layers was trained using the Adam optimizer with early stopping. Cobb angles were estimated from the predicted landmark coordinates using a fixed geometric post-processing rule. Model performance was evaluated against the clinically measured reference Cobb angle using mean absolute error (MAE), root mean square error (RMSE), threshold accuracy, and Bland-Altman analysis, and reproducibility was assessed by retraining the final model across three independent random seeds. The final model achieved a landmark localization MAE of 0.0111 ± 0.0005 and an RMSE of 0.0149 ± 0.0005 in normalized coordinate space (mean ± SD across three independently trained models). During preparation of this revision, an anisotropic coordinate-normalization issue was identified and corrected in the Cobb angle computation pipeline (see Methods); after this correction, the model achieved a Cobb angle MAE of 8.10° ± 0.38° against the rule-based GT and 7.86° ± 0.78° against the clinically measured reference (RMSE 10.90° ± 1.44° and 10.39° ± 1.22°, respectively; mean ± SD across three independently trained models). Threshold analysis showed that 25.4 ± 4.5% of cases were within 3°, 38.4 ± 2.5% were within 5°, and 71.7 ± 0.0% were within 10° against the rule-based GT, and 26.1 ± 9.5%, 41.3 ± 5.8%, and 70.3 ± 4.5%, respectively, against the clinically measured reference. Bland-Altman analysis (representative model, vs clinical reference) showed a mean bias of - 1.23°, with 95% limits of agreement from - 22.26° to 19.81°, reflecting a small number of higher-error cases among curves of varying severity. After correcting an anisotropic coordinate-normalization issue that had previously led to an underestimated Cobb angle and an optimistic performance estimate, the corrected evaluation showed that the proposed CNN-based coordinate regression model achieves a Cobb angle MAE against the clinically measured reference (7.86° ± 0.78°) within or near the range of inter-observer variability reported for manual Cobb angle measurement (6.2°-7.7°) [14, 26]. Repeated training across three independent random seeds showed a moderate degree of run-to-run variability (SD 0.78° for Cobb angle MAE against the clinically measured reference), underscoring the importance of evaluating reproducibility rather than relying on a single training run. These findings suggest that the method may be useful as an initial measurement-support tool, while further validation with larger and multi-center datasets is needed before clinical translation.