Sequence based local-global information fusion framework for vertebrae detection under pathological and FOV variation...

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Spinal Anatomy

Background:

• Automated vertebrae detection is crucial for spinal computer-aided systems.
• Challenges include unpredictable field-of-view and diverse pathological variations in vertebral morphology.

Purpose of the Study:

• To propose an effective sequence-based framework for robust and accurate vertebrae identification and localization.
• To overcome challenges posed by field-of-view variations and pathological cases in medical imaging.

Main Methods:

• A three-module framework: Local Feature Extraction (LFE), Discriminative Sequential Image Description (DSID), and Spinal Pattern Exploitation (SPE).
• LFE uses a shape-compatible sampler and CNN for local features and score maps.
• DSID constructs reliable vertebral feature sequences, preventing false positives/negatives.
• SPE fuses hierarchical local-global information using end-balanced relative position learning.

Main Results:

• Achieved an identification rate of 0.974 ±0.025 on a dataset of 450 spinal MRIs.
• Demonstrated a low localization error of 4.742 ±2.928 pixels.
• Showcased superiority over existing state-of-the-art methods in handling variations.

Conclusions:

• The proposed framework is effective for vertebrae detection, robust to field-of-view and pathological variations.
• It offers accurate identification and localization, outperforming current methods.
• This approach enhances the reliability of spinal computer-aided systems.