Introduction
Accurate prostate segmentation in Computed Tomography (CT) scans is a critical component of radiotherapy planning. However, CT images are characterized by low soft-tissue contrast, which often obscures the boundaries between the prostate and adjacent organs like the bladder or rectum. Traditional deep learning architectures, such as the standard U-Net, typically rely on pixel-wise loss functions (e.g., Cross-Entropy or Dice loss). These functions treat voxels as independent entities, failing to account for the semantic correlations between neighboring voxels, which often results in fragmented or anatomically inconsistent segmentations.
Methodology: The MetricUNet Framework
To overcome these limitations, we propose MetricUNet, a two-stage framework that integrates online voxel-level metric learning into the segmentation process.
Two-Stage Pipeline
- Coarse Localization: A lightweight U-Net first identifies the general prostate region from a down-sampled CT image to define a focused Region of Interest (ROI).
- Fine Segmentation: The core MetricUNet architecture employs a dual-branch design. While one branch produces the final segmentation mask, the second branch—the Voxel-Metric Learning Sub-network—supervises intermediate feature maps using metric losses (triplet and contrastive loss). This forces the network to cluster "prostate" features together and push "non-prostate" features away in the high-dimensional embedding space.
Online Voxel Sampling
Performing metric learning on every voxel in a 3D volume is computationally prohibitive. We introduced three online sampling strategies to make this process efficient:
- Focal Hard Negative Sampling: Targets voxels where the network is currently uncertain or incorrect, prioritizing learning from difficult regions.
- Contour-Aware Sampling: Specifically samples voxels near the prostate boundaries, where segmentation errors are most likely to occur.
- MetricUNet-HCP: The most advanced variant, combining both Hard negative and Contour-aware sampling.
Results and Discussion
The MetricUNet-HCP model was evaluated on a primary dataset of 339 clinical CT scans and externally validated on the PROMISE 2012 MRI dataset.
As illustrated in the figure above, the inclusion of metric learning significantly enhances the discriminative power of the feature space. Quantitatively, MetricUNet-HCP achieved a Dice Similarity Coefficient (DSC) of 0.8839 on the CT dataset, a 4.6% improvement over the baseline U-Net (0.8376). Boundary precision also improved dramatically, with the Average Surface Distance (ASD) dropping from 3.89mm to 1.39mm. Furthermore, the model outperformed the industry-standard nnU-Net across all major metrics with statistical significance ($p < 0.05$).
Conclusion
By regularizing the feature space through online voxel-metric learning, MetricUNet effectively addresses the challenges of low-contrast CT imaging. This multi-task approach produces smoother, more anatomically plausible segmentations with minimal computational overhead. Future research will explore single-stage end-to-end architectures and semi-supervised techniques to further refine these results and reduce the need for large-scale manual delineations.
Summary updated