Source-linked AI summary
Efficient Semi-Supervised Gross Target Volume of Nasopharyngeal Carcinoma Segmentation via Uncertainty Rectified Pyramid Consistency
Xiangde Luo, Wenjun Liao, Jieneng Chen, Tao Song, Yinan Chen, Shichuan Zhang, Nianyong Chen, Guotai Wang, Shaoting Zhang
TL;DR
NPC GTV segmentation needs accurate contours for radiotherapy, but obtaining the labeled images required by deep networks is expensive and time-consuming. URPC uses pyramid predictions, multi-scale consistency, and single-pass uncertainty rectification to exploit unlabeled data. It improved segmentation with limited labels, outperformed five semi-supervised methods, and reached an average Dice score of 82.74% with 50% labeled images.
Problem
Accurate NPC GTV delineation is important for radiotherapy, but deep-learning segmentation depends on expensive, time-consuming labeled images.
Method
URPC combines pyramid predictions with multi-scale consistency for unlabeled images and rectifies uncertainty estimated from prediction discrepancies in a single forward pass.
Results
URPC improved segmentation with limited labels, outperformed five semi-supervised methods using 10% labeled data, and achieved an average Dice score of 82.74% with 50% labeled images.
Takeaways & Limitations
URPC demonstrated effectiveness and generalization for GTVnx and GTVnd segmentation and showed potential for further clinical use.
Abstract
from arXiv · showhide
Gross Target Volume (GTV) segmentation plays an irreplaceable role in radiotherapy planning for Nasopharyngeal Carcinoma (NPC). Despite that Convolutional Neural Networks (CNN) have achieved good performance for this task, they rely on a large set of labeled images for training, which is expensive and time-consuming to acquire. In this paper, we propose a novel framework with Uncertainty Rectified Pyramid Consistency (URPC) regularization for semi-supervised NPC GTV segmentation. Concretely, we extend a backbone segmentation network to produce pyramid predictions at different scales. The pyramid predictions network (PPNet) is supervised by the ground truth of labeled images and a multi-scale consistency loss for unlabeled images, motivated by the fact that prediction at different scales for the same input should be similar and consistent. However, due to the different resolution of these predictions, encouraging them to be consistent at each pixel directly has low robustness and may lose some fine details. To address this problem, we further design a novel uncertainty rectifying module to enable the framework to gradually learn from meaningful and reliable consensual regions at different scales. Experimental results on a dataset with 258 NPC MR images showed that with only 10% or 20% images labeled, our method largely improved the segmentation performance by leveraging the unlabeled images, and it also outperformed five state-of-the-art semi-supervised segmentation methods. Moreover, when only 50% images labeled, URPC achieved an average Dice score of 82.74% that was close to fully supervised learning.
1 Introduction
NPC GTV delineation is important for radiotherapy but manual annotation is laborious, while deep-learning segmentation requires costly labeled data. URPC addresses this semi-supervised setting by combining multi-scale consistency with uncertainty-based stabilization.
- Manual contouring of GTVnx and GTVnd is tedious and laborious, while accurate annotations require substantial time and expertise.
- Semi-supervised learning can reduce annotation efforts by exploiting unlabeled medical images through regularization or pseudo-labeling.
- URPC uses pyramid predictions and encourages consistency across multiple scales, alongside supervised multi-scale learning on labeled images.
- An uncertainty rectifying module emphasizes reliable low-uncertainty predictions and weakens unreliable high-uncertainty predictions using discrepancy among scales from a single forward pass.
2 Methods
URPC combines pyramid predictions, supervised multi-scale learning, and uncertainty-rectified consistency to exploit labeled and unlabeled data for semi-supervised segmentation.
- Multi-Scale Prediction Network: PPNet adds prediction layers across decoder upsampling blocks to produce multi-scale segmentation outputs from a 3D UNet backbone.Each prediction layer uses 1×1×1 convolution followed by softmax; dropout and feature-level noise are inserted before prediction layers.
- Supervised Learning: Labeled images train PPNet with a multi-scale combination of Dice and cross-entropy losses.The supervised loss is applied to predictions at multiple scales against the ground-truth labels.
- Pyramid Consistency: Unlabeled images receive pyramid consistency regularization that minimizes prediction variance through L2 distances from the cross-scale average prediction.The predictions are first rescaled to the input size before consistency is encouraged.
- Uncertainty Rectification: Different spatial resolutions produce different frequency content, making direct voxel-level consistency prone to lost fine details or model collapse.Lower-resolution predictions capture low-frequency segmentation components, whereas higher-resolution predictions retain more high-frequency components.
- Uncertainty Rectification: URPC estimates scale-level uncertainty from prediction discrepancies in one forward pass, then weights reliable voxels more heavily and minimizes uncertainty during unsupervised training.KL-divergence measures uncertainty relative to the average prediction; higher uncertainty yields a lower voxel-wise weight, while uncertainty minimization promotes cross-scale agreement.
- Overall Objective: The overall objective combines supervised and uncertainty-rectified unsupervised losses with a time-dependent Gaussian warm-up controlling their balance.The weighting function increases the influence of unsupervised consistency according to training progress.
3 Experiments and Results
Experiments evaluated URPC on the NPC MRI dataset through implementation details, ablations, and comparisons with supervised and semi-supervised baselines.
- Dataset: The dataset comprised 258 T1-weighted MRI images, split into 180 training, 20 validation, and 58 testing cases.Training used 18 labeled images and 162 unlabeled images in the 10% labeled setting.
- Implementation: The implementation used PyTorch with SGD optimization, a batch size of four, and batches containing two annotated and two unannotated images.Training used eight TITAN 1080TI GPUs and a poly learning-rate schedule.
- Ablation study: Using four pyramid scales produced the best ablation setting, while five scales reduced performance because the lowest-resolution prediction lost details.Both uncertainty rectification and uncertainty minimization improved performance, and combining all submodules produced the strongest gain.
- Comparison with baselines: With 10% labeled data, URPC outperformed five state-of-the-art semi-supervised methods and required less training time and computational cost.Compared methods included MT, ICT, EM, UAMT, and DAN; DAN was the strongest existing method.
4 Conclusion
The conclusion presents URPC as an efficient semi-supervised framework for medical image segmentation, using pyramid consistency and uncertainty rectification to learn from unlabeled images.
- Conclusion: URPC combines multi-scale pyramid predictions with uncertainty rectification to stabilize learning from unlabeled images.Its uncertainty estimation requires only a single forward pass.
- Conclusion: Applied to GTVnx and GTVnd segmentation, URPC demonstrated effectiveness and generalization with potential for further clinical use.The authors identify evaluation on other segmentation tasks as future work.
Details of NPC dataset and methods’ implementations
The study used a curated NPC MRI dataset and evaluated methods under a 20% labeled-data comparison setting.
- Dataset: The NPC dataset was collected from West China Hospital and included newly diagnosed patients whose MRI examinations underwent strict inclusion and exclusion criteria.The supplied passage identifies the collection site and review period but does not provide the full criteria.
- Dataset: The dataset included clinical and tumor characteristics of the collected MRI-NPC cases, as summarized in Fig. 3.The supplied figure caption does not specify the individual characteristics shown.
- Evaluation setting: Table 3 compares URPC with existing methods on the NPC MRI dataset using 20% labeled data, with asterisks indicating p-value < 0.05 comparisons.The supplied table caption does not provide the numerical comparison values.
Results on 20% labeled NPC data.
With 20% labeled data, URPC achieved the best performance among the compared semi-supervised methods and produced more complete visual segmentations with uncertainty concentrated near boundaries.
- Quantitative results: URPC achieved the best performance over the state-of-the-art semi-supervised methods when using 20% labeled data.All semi-supervised methods improved substantially over supervised learning in this setting.
- Visual results: URPC showed higher overlap with ground truth and reduced false negatives compared with supervised learning and EM, especially for GTVnd segmentation.The comparison was reported at both slice and volume levels.
- Uncertainty analysis: The estimated uncertainty was mainly distributed near segmentation boundaries.The visualization used 20% labeled and 80% unlabeled data, with lime and yellow contours denoting predictions and ground truth.
Discussion and future work
URPC improves semi-supervised learning efficiency by estimating uncertainty from multi-scale prediction variance in one forward pass and automatically selecting reliable pixels. The framework is presented as task-agnostic beyond NPC GTV segmentation.
- Discussion and future work: URPC estimates uncertainty from the variance among PPNet’s multi-scale predictions in a single forward pass.This avoids the multiple forward passes required by commonly used MC dropout.
- Discussion and future work: The method automatically selects reliable pixels instead of relying on a manually designed uncertainty threshold.
- Discussion and future work: URPC is applied to NPC GTV segmentation but does not rely on task-specific knowledge, supporting use for other lesions and organs.