Two-Stream Compare and Contrast Network for Vertebral Compression Fracture Diagnosis

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Differentiating benign (osteoporotic/traumatic) and malignant vertebral compression fractures (VCFs) is crucial for patient treatment.
• Current automated VCFs diagnosis uses a two-step approach (identification then classification).
• Challenges include high intra-class variation, high inter-class similarity, and extreme class imbalance in VCF datasets.

Purpose of the Study:

• To develop a novel, integrated approach for VCFs diagnosis.
• To model VCFs diagnosis as a single three-class classification problem (normal, benign VCF, malignant VCF).
• To address the limitations of existing two-step methods and dataset challenges.

Main Methods:

• Proposed a novel Two-Stream Compare and Contrast Network (TSCCN).
• The network features a recognition stream for VCF identification by comparing adjacent vertebrae.
• Incorporates a classification stream for fine-grained differentiation using intra- and inter-class comparisons, integrated by a learnable weight control module.

Main Results:

• Achieved high diagnostic performance on a dataset of 239 VCFs patients.
• Demonstrated average sensitivity of 92.56% and specificity of 96.29% for VCFs diagnosis.
• Successfully treated VCFs diagnosis as a three-class problem, overcoming previous limitations.

Conclusions:

• The TSCCN offers an effective and accurate method for VCFs diagnosis.
• This integrated, three-class approach shows promise for improving automated VCFs detection and classification.
• The network's design addresses key challenges in VCF recognition and classification.