Poster Presentation 50th International Society for the Study of the Lumbar Spine Annual Meeting 2024

Introduction:

Adolescent Idiopathic Scoliosis (AIS) is the most common spinal deformity that affects children. Among the numerous parameters utilized to measure spinal deformities of AIS patients, the thoracic kyphosis (TK) angle and lumbar lordosis (LL) angle are essential for spine balance assessment. Monitoring spinal alignment involves physical examinations and subsequent X-ray imaging. However, physical examinations can be subjective, while X-ray imaging exposes patients to harmful radiation and with restricted access. We have developed and validated an innovative portable system and device utilizing 3D imaging and deep learning for coronal alignment analysis. This study aims to further develop the sagittal alignment analysis with internal validations.

Method:

This study enrolled consecutive 1671 AIS patients who underwent lateral X-ray scans at the Duchess of Kent Children's Hospital at Sandy Bay. 3D images of the unclothed back were obtained with whole spine radiographs taken on the same day. Our deep learning model utilized the RGBD images as input and generated radiologically comparable images (RCI), with real X-rays serving as the Ground truth (GT). Image similarity index, i.e., structural similarity between the RCI and the corresponding real X-ray was calculated. Further, the TK and LL computed from the synthesized and real radiographs using our previously developed online platform mskalign® were confirmed and compared by two spine specialists using linear regressions. Deformity severity classifications were compared using confusion matrix with the sensitivity and specificity reported.

Results:

The obtained RCIs were evaluated in terms of structural similarity index (SSIM=0.612). The TK and LL angles, derived from the synthesized RCI, demonstrate a strong correlation with the GT angle (TK: R2 = 0.827, p < 0.001; LL: R2 = 0.735, p < 0.001). The sensitivity and specificity of the RCI for identifying abnormal TK are 0.83 and 0.91, respectively. The sensitivity and specificity of the RCI for identifying abnormal LL are 0.80 and 0.90, respectively.

Discussion:

The RCIs exhibit high similarity to the real X-rays, and the TK and LL measured on the RCIs demonstrate strong consistency with the GT obtained from the real X-rays. Although we propose that our new device and system offer accurate, radiation-free, and accessible spinal alignment analysis, further independent validations are required.

Figure 1. Confusion matrix for identifying abnormal TK and LL. (a) Confusion matrix of TK. (b) confusion matrix of LL.