Scoliosis screening system based on improved optical photogrammetry using Kinect sensor

Technical Summary: Scoliosis Screening System Based on Improved Optical Photogrammetry Using Kinect Sensor

Problem Statement
Scoliosis is a prevalent spinal disorder among adolescents, with a global prevalence of 0.5–5.2%. Early detection is critical to prevent severe thoracic deformity and visceral dysfunction. Current screening methods face significant limitations: manual physical examinations are inefficient for large populations and rely heavily on physician experience, while X-ray imaging, the clinical gold standard (Cobb angle), involves radiation exposure, high costs, and complex procedures, making it unsuitable for routine, frequent monitoring in schools or communities. Existing non-invasive optical methods (e.g., Moiré topography, laser scanners) often require expensive hardware, strict lighting conditions, or complex calibration. Furthermore, while deep learning approaches have emerged, they often lack clinical interpretability and require massive annotated datasets. There is a need for a low-cost, radiation-free, portable, and efficient screening system that maintains diagnostic accuracy comparable to X-rays.

Methodology
The authors developed a scoliosis screening system utilizing a Microsoft Kinect sensor to capture color and depth images, which are aligned to generate a 3D point cloud of the subject's back. The methodology proceeds through the following stages:

1. Data Acquisition: Subjects stand in two positions: vertically (for coronal plane analysis) and in a 90° forward bend (for transverse plane analysis). The system operates within a 0.8–2.5m range.
2. Point Cloud Processing:
   • Filtering: Pass-through and outlier filters remove background and noise points.
   • Segmentation: A color-based region-growing algorithm (converting RGB to HSV for better landmark distinction) isolates the back and upper limbs from hair and clothing. The study determined that a color distance threshold of 5 provided the optimal balance between segmentation accuracy and processing time.
   • Smoothing: A moving least squares method (polynomial order 2, sub-region radius 0.02m) is applied to reduce measurement errors while maintaining efficiency.
3. Feature Extraction:
   • Shoulder Height: Edge points of the shoulders are fitted with cubic polynomials to determine height differences.
   • Anatomical Landmarks: Specific markers (C7, L5, and ATR measurement points A and B) are manually placed on the subject and extracted via HSV color thresholding.
   • Spine Curve Fitting: Concave points on the back are identified by comparing Z-values of neighboring points within Y-axis fragments (fragment length 0.002m, concave coefficient 0.9). These points are projected onto coronal and sagittal planes and fitted using the least squares method.
4. Multi-Index Evaluation: The system calculates four metrics to assess spinal health across three anatomical planes:
   • Uneven Shoulders: Height difference and angle between shoulders.
   • DOSA Angle (Deep Optics-based Scoliosis Assessment): A coronal plane curvature angle calculated by deriving the spine's fitted curve to find points of maximum curvature change, analogous to the Cobb angle but derived from surface data.
   • Kyphosis Index (Sinoto Value): A sagittal plane metric representing the ratio of hump height to back length.
   • Angle of Trunk Rotation (ATR): A transverse plane metric calculated from the height difference and horizontal distance between specific points during the forward bend.

Key Results
The system was validated on 40 subjects (ages 7–77) with suspected scoliosis, comparing DOSA angles against X-ray-derived Cobb angles.

• Correlation: The mean difference between DOSA and Cobb angles was -0.386° with a variance of 7.773. 87.5% of the differences fell within ±3°. Statistical analysis (t-test, p=0.05) showed no significant difference between the two methods (t = -0.877).
• Diagnostic Performance: Using a 10° threshold for scoliosis, the system achieved a sensitivity of 83.33%, specificity of 90.00%, and overall accuracy of 85.00%. The Area Under the Curve (AUC) for the ROC analysis was 0.967.
• ATR Accuracy: The system demonstrated 100% consistency with control measurements for identifying abnormal trunk rotation, with a maximum error of 0.408°, significantly outperforming the typical 2°–3° variability of manual Scoliometer readings.
• Kyphosis Analysis: The Sinoto value indicated a higher prevalence of kyphosis in subjects over 50, showing little correlation with the DOSA angle, confirming the system's ability to distinguish between lateral and sagittal deformities.

Significance and Claims
The paper claims that the proposed system offers a viable alternative for primary scoliosis screening in community and school settings. Its significance lies in:

• Accessibility: It utilizes a portable, low-cost Kinect sensor, eliminating radiation risks and the need for expensive laser systems or controlled lighting environments.
• Efficiency: The entire process, including operator learning (approx. 5 minutes) and measurement (approx. 2 minutes), is streamlined within a single software program.
• Comprehensive Assessment: Unlike methods focusing solely on the coronal plane, this system provides multidimensional data (coronal, sagittal, and transverse) including shoulder asymmetry, kyphosis, and trunk rotation, offering a more holistic view of spinal deformity.
• Clinical Consistency: The DOSA angle demonstrates strong agreement with the traditional Cobb angle, suggesting it can serve as a reliable non-invasive proxy for X-ray assessment in screening scenarios.

Limitations
The authors acknowledge that the system's accuracy may decrease for individuals with a BMI > 28 kg/m² or excessive loose skin, where the spinal concavity is less distinct. Additionally, the study notes that a larger sample size is required to further validate measurement errors and generalizability.