VLTENet: A Deep-Learning-Based Vertebra Localization and Tilt Estimation Network for Automatic Cobb Angle Estimation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Scoliosis diagnosis relies on Cobb angle estimation from spinal X-rays.
• Current deep learning methods for automated scoliosis assessment often use regression models lacking structural information.
• Landmark detection and vertebra segmentation approaches are sensitive to errors.

Purpose of the Study:

• To propose a novel deep learning architecture, VLTENet, for accurate Cobb angle estimation in scoliosis.
• To improve automated scoliosis assessment by directly predicting vertebra localization and tilt.
• To enhance accuracy by integrating structural spine information into the deep learning model.

Main Methods:

• Developed VLTENet, combining HRNet and U-Net architectures for comprehensive feature extraction.
• Incorporated a Feature Fusion Channel Attention (FFCA) module to prioritize informative features.
• Introduced a Joint Spine Loss (JS-Loss) function to focus on spinal regions and constraints.
• Proposed a new Cobb angle estimation method aligned with clinical guidelines.

Main Results:

• VLTENet demonstrated superior performance in Cobb angle estimation compared to existing methods.
• The model achieved accurate estimates for various scoliosis types.
• Experiments on public and in-house datasets validated the method's effectiveness.

Conclusions:

• VLTENet offers a significant advancement in automated scoliosis assessment.
• The proposed architecture and loss function effectively improve Cobb angle estimation accuracy.
• This method holds promise for more reliable and precise scoliosis diagnosis.