Two-stage video-based convolutional neural networks for adult spinal deformity classification

Area of Science:

• Biomedical Engineering
• Computer Science
• Orthopedics

Background:

• Current methods for diagnosing spinal disorders, including adult spinal deformity (ASD), rely on static assessments or dynamic observations that are labor-intensive or facility-dependent.
• There is a need for a diagnostic approach that is facility-independent, has low practice flow, and does not involve patient contact.

Purpose of the Study:

• To develop and evaluate a video-based, two-stage machine learning method for classifying patients with spinal disorders, specifically ASD and other forms.
• To create a diagnostic tool that overcomes the limitations of existing static and dynamic assessment methods.

Main Methods:

• A two-stage machine learning approach was employed, utilizing deep learning for patient localization and a 3D convolutional neural network (CNN) for extracting spatial-temporal gait information from videos.
• The model was trained and validated on a dataset of 81 patients, assessing performance using mean accuracy, F1 score, and AUROC, with five-fold cross-validation.

Main Results:

• The proposed video-based method achieved a mean accuracy of 0.7553, an F1 score of 0.7063, and an AUROC of 0.7864 in classifying spinal disorders.
• Ablation experiments confirmed the significance of the detection stage and transfer learning in the model's performance.
• The method demonstrated superior performance compared to professional observations of gait videos, achieving higher accuracy and AUROC scores.

Conclusions:

• The developed video-based machine learning framework offers an effective and reliable method for diagnosing ASD and other spinal disorders.
• This approach has the potential to improve clinical diagnosis by reducing facility dependency and providing data-driven insights, benefiting both patients and clinicians.