Deep learning-based rotational object detection algorithm for automatic Cobb angle measurement in X-ray images of scoliosis.

BACKGROUND

Scoliosis, characterized by an abnormal curvature of the spine, poses significant physiological challenges during adolescent growth and development. The Cobb angle measurement is regarded as the clinical gold standard for assessing the severity of scoliosis. The measurement of the Cobb angle has been extensively investigated through advanced neural network models; however, accurately measuring the Cobb angle remains challenging due to difficulties in detecting rotational vertebral objects. Therefore, this study aimed to develop an enhanced model that improves the detection of rotational vertebrae, thereby increasing the accuracy of Cobb angle measurements for more reliable scoliosis assessment.

METHODS

This study employed a redesigned YOLOv8-DSF model based on the YOLOv8n-oriented bounding box (OBB) framework, equipped with innovative modules to improve the detection of rotational objects. The method involves detecting vertebrae with rotational bounding boxes using the model and subsequently calculating the Cobb angle based on these detections. The model was tested and evaluated on a custom-built dataset that integrates private and publicly available data, all of which were enhanced through rigorous screening and augmentation processes.

RESULTS

The experimental findings demonstrated that the model achieved an mAP calculated with an Intersection over Union (IoU) threshold of 0.50 (mAP50) score of 0.626 and an mAP calculated with an IoU range of 0.50 to 95 (mAP50-95) score of 0.424, both exceeding the baseline by over 10%. In assessing Cobb angle accuracy, the model achieves a symmetric mean absolute percentage error (SMAPE) of 8.43 and an average mean absolute error (MAE) of 5.09 across the three Cobb angles. When tested on a public dataset for Cobb angle prediction accuracy, YOLOv8-DSF also demonstrated excellent performance.

CONCLUSIONS

Given that the model outperformed both the baseline and other existing methods in detecting rotational vertebral objects and calculating the Cobb angle, it contributes to more accurate scoliosis assessments. This advancement holds significant potential for clinical applications in scoliosis evaluation and management.