A Dual-Feature Extractor Framework for Accurate Back Depth and Spine Morphology Estimation from Monocular RGB Images.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Adolescent idiopathic scoliosis (AIS) assessment relies on X-rays, posing risks like radiation exposure and limited access.
• Current RGB image analysis for scoliosis is unstable due to environmental factors like lighting.
• A need exists for accessible, radiation-free scoliosis assessment methods.

Purpose of the Study:

• To develop a novel pipeline for accurate spine morphology estimation using depth and surface information from unclothed back images.
• To compensate for the limitations of 2D imaging in scoliosis assessment.
• To improve the stability and generalizability of scoliosis analysis models.

Main Methods:

• Proposed an adaptive multiscale feature learning network, GAMA-Net, for precise depth estimation of the back surface.
• Utilized dual encoders for patch-level and global feature extraction, with a Patch-Based Hybrid Attention (PBHA) module.
• Employed an Adaptive Multiscale Feature Fusion (AMFF) module for dynamic information fusion in the decoder.

Main Results:

• The GAMA-Net model achieved high accuracy in depth estimation, with scores of 78.2%, 93.6%, and 97.5% across three metrics.
• Integrating depth and surface information for spine morphology estimation resulted in up to 97% accuracy in spine curve generation.
• The proposed approach demonstrated significant improvements in accuracy and reliability for scoliosis assessment.

Conclusions:

• The novel pipeline effectively estimates depth information, enhancing spine morphology analysis for scoliosis.
• This integrated approach offers a promising, radiation-free alternative to traditional X-ray assessments for AIS.
• The developed GAMA-Net model shows potential for clinical relevance in scoliosis diagnosis and monitoring.