Comprehensive Summary
This study introduces the Dual AttentionUNet, a deep learning model designed for the efficient screening and diagnosis of scoliosis using bare-back images. The research involved 458 participants, including 350 scoliosis patients and 108 healthy controls, categorized based on Cobb angle measurements into healthy, mild, moderate, and severe groups. The model enhances the traditional U-Net architecture by incorporating channel and self-attention mechanisms, improving feature extraction in medical images. It segments the back contour and computes an asymmetry index correlating with scoliosis severity. The model achieved an overall accuracy of 86.9%, with precision and recall rates of 87.2% and 86.9%, respectively. AUC scores ranged from 0.874 for mild cases to 0.954 for severe cases, demonstrating performance comparable to or better than experienced clinicians. The model also provided faster evaluations than human experts, highlighting its efficiency in large-scale screening.
Outcomes and Implications
The Dual AttentionUNet model offers a significant advancement in scoliosis screening by providing a non-invasive, cost-effective, and rapid alternative to traditional methods. It reduces the need for time-consuming physical examinations and radiation-heavy imaging, making it suitable for large-scale screenings. The model's high accuracy and speed can minimize diagnostic delays and unnecessary X-rays, democratizing scoliosis screening globally. However, its performance is limited in patients with higher BMI or subtle deformities, and the relatively small dataset may restrict generalizability. Future improvements could involve larger, diverse datasets and mobile-based deployment for rural access, enhancing its clinical relevance and impact.