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

Abstract

Background Spinal tuberculosis is the most common extrapulmonary manifestation of tuberculosis. Magnetic resonance imaging (MRI), particularly fat-suppressed T2-weighted imaging (FS-T2WI), is the modality of choice for preoperative evaluation; yet indistinct lesion boundaries render manual delineation subjective and poorly reproducible. Although deep learning segmentation has advanced considerably, its “black-box” nature and limited interpretability remain critical obstacles. This study aimed to develop an uncertainty-guided deep learning segmentation framework and evaluate its accuracy and clinical utility. Methods We retrospectively enrolled 210 patients with spinal tuberculosis from an initial cohort of 300 screened at three centers, and acquired preoperative FS-T2WI scans. Data from Center 1 (n = 160) were used for five-fold cross-validation, while the remaining 50 cases served as an external test set. We built an improved model on nnU-Net by integrating boundary-aware loss with Monte Carlo Dropout, and compared it against U-Net, Attention U-Net, and TransUNet. The uncertainty threshold was determined through internal cross-validation, and clinical validation followed a within-subject crossover design involving nine physicians. Results On the external test set, the improved model achieved a Dice similarity coefficient of 0.858 and an AUC of 0.912, outperforming all comparative models. Uncertainty correlated strongly and positively with pixel-level error rates (Spearman ρ = 0.74). At a threshold of 0.52, sensitivity for identifying unreliable segmentation was 84.3% and negative predictive value was 89.5%. In the physician validation, overlaying uncertainty maps significantly increased trust scores (3.8 ± 0.7 vs. 3.2 ± 0.8; P < 0.001, Cohen’s d = 0.80) and reduced review time (P = 0.018), with the greatest benefit observed among resident physicians. In cases of obvious AI failure, uncertainty alerts accurately flagged erroneous regions in 80.0% of instances. Conclusion The proposed uncertainty-guided framework improved spinal tuberculosis lesion segmentation accuracy. Pixel-level uncertainty maps reliably identified unreliable predictions and enhanced clinician trust, offering a robust AI decision-support tool for precise surgical planning.