EDRNet-based adolescent idiopathic scoliosis screening method using bare-back images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Orthopedics

Background:

• Adolescent idiopathic scoliosis (AIS) affects 2-3% of adolescents.
• Current X-ray screening methods for AIS pose limitations due to radiation exposure and equipment costs.
• There is a need for radiation-free and efficient screening tools for AIS.

Purpose of the Study:

• To develop a novel, radiation-free screening method for AIS using bare-back images.
• To evaluate the performance of an efficient dilated residual network (EDRNet) for scoliosis classification.
• To compare the diagnostic capabilities of EDRNet with human clinicians.

Main Methods:

• A dataset of bare-back images from 600 adolescent patients (300 with AIS, 300 without) was utilized.
• Data augmentation techniques were applied to expand the dataset.
• Five deep learning models (VGG16, GoogLeNet, ResNet34, MobileNetV2, EDRNet) were trained and compared for scoliosis classification.
• Performance was assessed using accuracy, precision, recall, F1 score, and ROC curves.

Main Results:

• All five deep learning models showed good predictive performance.
• The proposed EDRNet model achieved the highest predictive performance among all tested models.
• EDRNet demonstrated a competitive classification ability (AUC=0.91) compared to clinicians (AUC=0.86-0.93) and offered superior speed.

Conclusions:

• The EDRNet model shows significant potential for accurate and efficient scoliosis classification using bare-back images.
• This radiation-free approach offers a promising solution for large-scale adolescent scoliosis screening.
• EDRNet can serve as a valuable tool to aid in early scoliosis detection and management.