Contrastive uncertainty learning for robust multi-structure segmentation in lower-pelvic MR
Prostate MRI segmentation suffers from regional boundary ambiguity, causing standard global-overlap losses to overlook clinically significant surface deviations. This paper proposes Contrastive Uncertainty Learning (CUL), which treats predictive uncertainty as a representation-level supervisory signal rather than a post-hoc filter: low-uncertainty same-class voxels are clustered in embedding space while high-uncertainty or misclassified voxels are actively repelled. Evaluated on 589 lower-pelvic MR scans across eight anatomical structures, CUL preserves mean Dice while significantly improving surface-sensitive metrics and suppressing extreme errors in hard cases, suggesting practical value for radiotherapy planning and contour-acceptance workflows.
CUL is a conceptually clean and clinically motivated supervised contrastive framework that targets geometric robustness rather than volumetric overlap. The paper supports its central claim with a carefully designed ablation hierarchy and a compelling hard-case asymmetry analysis, though limited test-set size ($n=117$), an unexplained fixed threshold $\tau=0.5$, and the absence of cross-dataset validation temper the generality of the conclusions.
The ablation study is the paper’s clearest strength. Replacing uncertainty guidance with class-based sampling and the contrastive term with a boundary loss yields a stepwise degradation from 6/8 to 5/8 to 4/8 significant structures, allowing precise attribution of each component’s contribution. The case-level asymmetry—18.8% significantly improved versus 0.9% significantly degraded in surface distance—provides credible evidence that CUL acts as geometric insurance against extreme boundary outliers.
The fixed normalised uncertainty threshold $\tau=0.5$ is stated without justification or sensitivity analysis, yet the ablations identify uncertainty-guided selection as the decisive factor for the transition zone; this raises concerns about generalisation if the threshold is dataset-dependent.
The authors also highlight a descriptively halved HD95 standard deviation (8.14 mm to 4.88 mm) but simultaneously report that Brown-Forsythe testing ($p=0.538$) fails to confirm a variance difference, creating tension between the narrative of outlier suppression and the inferential evidence. Moreover, evaluation is confined to a single dataset and a single ViT+UperNet backbone, leaving claims of broad applicability to multi-structure MRI segmentation under-substantiated.
The core evidence is strong: paired Wilcoxon signed-rank testing confirms a statistically significant reduction in mean surface distance ($p<0.0001$) across six of eight structures, whereas Dice and HD95 median shifts are not significant ($p=0.181$ and $p=0.816$, respectively). The authors fairly distinguish their fully supervised, embedding-structuring approach from prior semi-supervised methods that use uncertainty only for filtering pseudo-labels. However, the paper does not provide quantitative head-to-head comparisons against these related methods or against modern geometric losses, leaving unresolved whether a tuned alternative could match the gains without the training-time cost of Monte Carlo Dropout.
Reproduction is partially documented but hampered by missing implementation details. While the authors report training for 100 epochs on an NVIDIA A40 with Adam at learning rate $10^{-4}$, batch size 4, and $M=5$ MC samples, they omit dropout placement and rate, the architecture of the contrastive projection head, the number of negative samples drawn per anchor, and the initialisation or learned range of the temperature $T$. No code repository is referenced, the dataset is cited but not linked, and the selection of the fixed threshold $\tau=0.5$ is not accompanied by a search protocol, all of which would impede independent reproduction.
Boundary delineation of anatomical structures in prostate MRI is often compromised by regional ambiguity. Automated segmentation methods relying on global overlap loss functions tend to be insensitive to surface deviations, leading to significant errors in ambiguous regions. Most existing uncertainty-aware methods treat uncertainty as a post-hoc quality indicator or prediction filter. We propose Contrastive Uncertainty Learning (CUL), a supervised framework that uses predictive uncertainty to directly guide representation learning. Using Monte Carlo Dropout, CUL identifies voxel-wise uncertainty and structures the embedding geometry accordingly: low-uncertainty, correctly predicted features form stable voxels while high-uncertainty or misclassified features are actively repelled, preventing feature collapse at difficult boundaries. We evaluate CUL on a 589-scan 3D prostate MRI multi-structure segmentation task (8 anatomical structures) using a ViT with UperNet decoder architecture. Average Dice ($83.38\%$) remains comparable to the baseline ($83.35\%$), while surface distance decreases significantly ($p<0.001$) across 6 of 8 anatomical structures. A hard-case analysis reveals a $22{\times}$ asymmetry between significantly improved and degraded cases ($18.8\%$ vs $0.9\%$). Ablation studies show that removing uncertainty-guided sample selection eliminates improvements for the transition zone, while replacing the contrastive objective with a boundary loss additionally removes bladder improvements, demonstrating that both uncertainty-guided sampling and the contrastive objective are essential for the observed gains. These findings demonstrate that the proposed uncertainty-guided representation learning is a useful mechanism for reducing the risk of extreme errors in safety-critical medical segmentation
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.