Convolutional Neural Network-Based Approach for Cobb Angle Measurement Using Mask R-CNN

Scoliosis is a disorder characterised by an abnormal spinal curvature, which can lead to negative effects on patients, affecting their quality of life. Given its progressive nature, the classification of severity requires an accurate diagnosis and effective monitoring. The Cobb angle measurement method has been widely considered as the gold standard for a scoliosis assessment. Commonly, an expert assesses scoliosis severity manually by identifying the most tilted vertebrae of the spine. However, this method is tedious, time-consuming, and presents limitations in measurement accuracy due to the intraobserver and interobserver variability. This highlights the need for a more objective tool less sensitive to manual intervention. Nowadays, advancements in artificial intelligence are transforming the diagnosis of scoliosis. In this study, we propose a fully automated approach to measure the Cobb angle. A small dataset of 98 anterior-posterior full spine X-ray images was labelled and included for evaluation. We assessed the accuracy and performance of the Mask R-CNN architecture for spine detection and segmentation. Beyond the neural network´s performance, a workflow was developed to enable midline identification, detection of the most significantly tilted vertebrae, direct visualization of Cobb angles, and scoliosis severity assessment. The model achieved high segmentation accuracy, with mIoU of 0.8012 and mDSC of 0.8878, while maintaining a mean precision of 0.9145. The mean Cobb angle was 25.43° ± 10.85° (range: 11.50-54.00°) for manual measurements by observer A, 25.89° ± 10.00° (range: 10.00-53.00°) by observer B, and 26.69° ± 12.50° (range: 10.29-59.34°) for automated measurements. We achieved a Mean Absolute Difference of 3.31º ± 2.69º, a Mean Absolute Error of 2.96° ± 2.60°, and an Intraclass Correlation Coefficient (95% CI) of 0.928 between manual and automated measurements. The automated method required an average of 3.3 seconds per radiograph. Although further improvements are needed, these results demonstrate the high potential of the proposed model, which provides experts with improved interpretability and precision in Cobb angle calculation and severity classification by overlaying them onto the original X-ray images.