Deep learning segmentation with uncertainty quantification for spinal tuberculosis on fat-suppressed T2-weighted MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Deep Learning for Medical Segmentation

Background:

• Spinal tuberculosis (TB) is the most common extrapulmonary TB manifestation.
• Magnetic resonance imaging (MRI), especially fat-suppressed T2-weighted imaging (FS-T2WI), is crucial for preoperative evaluation.
• Manual segmentation of spinal TB lesions on MRI is subjective and lacks reproducibility, hindering accurate surgical planning.

Purpose of the Study:

• To develop and evaluate an uncertainty-guided deep learning framework for segmenting spinal tuberculosis lesions on MRI.
• To assess the accuracy, interpretability, and clinical utility of the proposed AI framework.
• To improve the reliability of automated segmentation for enhanced surgical planning.

Main Methods:

• A retrospective study of 210 patients with spinal tuberculosis using preoperative FS-T2WI scans.
• Development of an improved nnU-Net model integrating boundary-aware loss and Monte Carlo Dropout.
• Comparison against U-Net, Attention U-Net, and TransUNet, with clinical validation involving nine physicians.

Main Results:

• The uncertainty-guided model achieved superior segmentation performance (Dice: 0.858, AUC: 0.912) on an external test set.
• Uncertainty maps strongly correlated with segmentation errors, enabling reliable identification of unreliable predictions (Sensitivity: 84.3%, NPV: 89.5%).
• Overlaying uncertainty maps significantly increased physician trust and reduced review time, particularly for residents.

Conclusions:

• The proposed uncertainty-guided framework enhances spinal tuberculosis lesion segmentation accuracy and interpretability.
• Pixel-level uncertainty maps serve as a reliable indicator of segmentation reliability.
• This AI decision-support tool offers robust capabilities for precise surgical planning in spinal TB cases.