Uncertain Boundaries, Uncertain Outcomes: Segmentation Disagreement Reveals Latent Risk In Locoregionally Advanced Nasopharyngeal Carcinoma
Abstract
Purpose
To evaluate the prognostic value of deep learning–derived spatial uncertainty in locoregionally advanced nasopharyngeal carcinoma (LA-NPC).
Methods
A retrospective internal cohort of 250 LA-NPC patients and an external validation cohort of 277 patients were analyzed. We employed the nnU-Net V2 framework integrated with Monte Carlo dropout (p=0.3) to estimate model uncertainty. For each patient, 30 stochastic forward passes were performed to generate probability maps. The "uncertain boundary" was strictly defined as the spatial union of voxels exhibiting fluctuating probabilities (0.1–0.9), distinguishing them from high-confidence tumor (>0.9) and background (9.7 mm) experienced significantly poorer outcomes, suggesting that high spatial uncertainty correlates with aggressive tumor behavior.
Conclusion
The ESST successfully translates abstract deep learning uncertainty into an interpretable clinical marker. Unlike traditional segmentation metrics, this uncertainty quantification captures latent risks associated with tumor boundary definitions. The high consistency with expert evaluation confirms its robustness, offering a novel, biologically relevant perspective for risk stratification in radiotherapy.