Automatic Measurement of Scoliosis Based on an Improved Segmentation Model

Abstract

The detection of the spine is crucial in automating the measurement of the Cobb Angle. While various segmentation models have been employed for vertebrae segmentation in X-ray images, there is a need to enhance segmentation performance. This paper proposes a comprehensive automatic measurement method for the Cobb angle. The RetinaNet model is employed to detect the region of interest corresponding to the spine, while the W-Net model is developed for accurate vertebrae segmentation. To address the issue of adjacent vertebrae adhesion in the segmented image, a post-processing technique is applied. Experimental results demonstrate that the W-Net model achieves superior performance, with a mean Intersection over Union (MIoU) of 0.9073 ± 0.0021, Dice Coefficient of 0.9446 ± 0.0139, and Precision of 0.9390 ± 0.0190. The post-processing step reduces adhesion at one end by approximately 83.4% and adhesion at both ends by approximately 83.6%. The reliability of the proposed method is evaluated through intra-group correlation coefficients (ICC) of 0.902 and 0.915, respectively, between two observers, both exceeding 0.9. The mean absolute deviation (MAD) is 3.08° and 2.91°, respectively. Therefore, the proposed method achieves automatic detection of the Cobb angle without the need for manual cropping or additional human intervention, while maintaining good reliability.

Keywords

• Cobb angle
• deep learning
• image segmentation
• scoliosis