Speaker
Description
Objective
In MR-guided adaptive radiotherapy (MRgRT) for prostate cancer, daily target re-delineation is required to account for inter-fraction anatomical changes, requiring rapid and reliable auto-segmentation. Image-quality degradation from acquisition- and patient-related factors can impair prostate MRI segmentation and may also affect uncertainty estimates. We systematically stress-tested structure-level uncertainty for prostate clinical target volume (CTV) segmentation under controlled MRI degradation, evaluating its response to degradation, association with segmentation quality, and quality control utility.
Methods
We used the LUND-PROBE pelvic T2-weighted MRI dataset, with 432 patients for model development and 35 patients as an independent hold-out test set, and a 3D nnU-Net segmentation model. Seven controlled perturbations were applied at six increasing severity levels: in-plane resolution, slice thickness, Gaussian noise, Rician noise, bias-field inhomogeneity, ghosting, and motion. These were designed as stress-test conditions representing clinically relevant degradation mechanisms rather than estimates of clinical prevalence. Uncertainty was estimated using test-time augmentation (TTA, $M=8$) and 10-fold cross-validation ensembles (CVE, $M=10$). Structure-level uncertainty was defined by spatially averaging voxel-wise predictive entropy or variance over the full image volume, predicted CTV, or a $\pm4$-pixel boundary band. Reliability was assessed using Spearman correlations with Dice similarity coefficient (DSC), normalized surface Dice (NSD), and 95th-percentile Hausdorff distance (HD95). QC utility was evaluated by detecting segmentation failure ($\mathrm{DSC}<0.85$) using AUROC and AUPRC.
Results
Structure-level uncertainty generally increased as segmentation performance deteriorated, but responses depended strongly on perturbation type and spatial aggregation. Associations with segmentation quality were weak on original images ($\rho_s=-0.01$ to $-0.46$) but strengthened under motion artifacts, reaching $\rho_s=-0.76$. Predicted-region uncertainty showed the most consistent negative associations with DSC, whereas global and boundary-band uncertainty showed stronger associations with NSD and HD95 in several perturbation settings. Predicted-region variance achieved AUROC/AUPRC values of 0.744/0.550 for TTA and 0.732/0.566 for CVE. All complete segmentation failures (DSC=0) had global uncertainty below the corresponding baseline P95 uncertainty threshold. TTA and CVE therefore showed similar overall failure-detection performance, although their uncertainty responses diverged under severe degradation.
Outlook
Structure-level UQ can support prostate MRgRT segmentation QC, but reliability depends on degradation type, spatial aggregation, and error type. Empty predictions should be detected explicitly before mask-dependent UQ is interpreted, while non-empty predictions should be complemented by surface-sensitive checks. Future work should validate these findings on naturally degraded MRI from external centers and different scanners.
Summary
Structure-level uncertainty can support QC of prostate CTV segmentation under prostate MRI degradation, but its reliability depends on degradation type and spatial aggregation.