Source-linked AI summary
Uncertainty-Guided Mutual Consistency Learning for Semi-Supervised Medical Image Segmentation
Yichi Zhang, Rushi Jiao, Qingcheng Liao, Dongyang Li, Jicong Zhang
TL;DR
Expert annotations for medical image segmentation are costly, and existing consistency methods do not fully use shape and boundary information from unlabeled data. The paper combines intra-task and cross-task consistency, guided by segmentation uncertainty, and reports improved results over supervised baselines and competing semi-supervised methods on two datasets.
Problem
Expert-annotated medical images are costly to acquire, and existing consistency-based methods do not fully exploit region-level shape constraints and boundary-level distance information from unlabeled cases.
Source: §1 · p. 2Method
The framework combines intra-task consistency from up-to-date predictions with cross-task consistency between segmentation and distance-map tasks, using teacher-estimated uncertainty to select relatively certain predictions.
Source: §1 · p. 2Results
The method improves Dice coefficient over supervised baselines by up to 4.13% on left atrium segmentation and 9.82% on brain tumor segmentation, and outperforms other semi-supervised methods on both datasets under the same backbone and task settings.
Source: §5 · p. 16Takeaways & Limitations
The authors identify potential for applying the method in further clinical applications to ease annotation costs.
Source: §5 · p. 16The stated methodology frames its objective around combining labeled and unlabeled cases and comparing performance with a model trained on fully labeled data.
Source: §3 · p. 5
Abstract
from arXiv · showhide
Medical image segmentation is a fundamental and critical step in many clinical approaches. Semi-supervised learning has been widely applied to medical image segmentation tasks since it alleviates the heavy burden of acquiring expert-examined annotations and takes the advantage of unlabeled data which is much easier to acquire. Although consistency learning has been proven to be an effective approach by enforcing an invariance of predictions under different distributions, existing approaches cannot make full use of region-level shape constraint and boundary-level distance information from unlabeled data. In this paper, we propose a novel uncertainty-guided mutual consistency learning framework to effectively exploit unlabeled data by integrating intra-task consistency learning from up-to-date predictions for self-ensembling and cross-task consistency learning from task-level regularization to exploit geometric shape information. The framework is guided by the estimated segmentation uncertainty of models to select out relatively certain predictions for consistency learning, so as to effectively exploit more reliable information from unlabeled data. Experiments on two publicly available benchmark datasets showed that: 1) Our proposed method can achieve significant performance improvement by leveraging unlabeled data, with up to 4.13% and 9.82% in Dice coefficient compared to supervised baseline on left atrium segmentation and brain tumor segmentation, respectively. 2) Compared with other semi-supervised segmentation methods, our proposed method achieve better segmentation performance under the same backbone network and task settings on both datasets, demonstrating the effectiveness and robustness of our method and potential transferability for other medical image segmentation tasks.
1. Introduction
Medical image segmentation needs extensive expert annotations, which are costly to obtain, while unlabeled scans are easier to acquire. The paper proposes uncertainty-guided mutual consistency learning to use unlabeled data through intra-task and cross-task consistency, with experiments reporting improved performance on two datasets.
1. Introduction
- Expert annotation is difficult and expensive, especially for 3D CT and MRI volumes; one 250-slice lung-infection CT scan takes experts about 400±45 minutes to delineate.Unlabeled medical images are relatively easier to obtain than expert-examined annotations.Source: §1 · p. 2
- Existing consistency-based methods do not fully exploit reliable region-level shape constraints and boundary-level distance information from unlabeled cases.Prior approaches include prediction consistency and task-level regularization for boundary-based surface mismatch.Source: §1 · p. 2
- The framework combines intra-task consistency from up-to-date predictions with cross-task consistency to exploit geometric shape information from unlabeled data.A dual-task backbone produces segmentation probability maps and signed distance maps; teacher-estimated uncertainty guides selection of relatively certain predictions for consistency learning.Source: §1 · pp. 2–3
- The method outperforms state-of-the-art semi-supervised segmentation methods on the left atrium and brain tumor datasets.The experiments use two publicly available medical image segmentation datasets.Source: §1 · p. 3
2. Related Work
Prior work spans segmentation architectures, semi-supervised learning strategies, and uncertainty estimation. This paper situates its approach among methods using CNNs, pseudo-labels or consistency learning, and uncertainty to select reliable information from unlabeled data.
2.1. CNNs for Medical Image Segmentation
- Encoder-decoder architectures, including U-Net and its 3D variants, are widely used for medical image segmentation; nnU-Net adapts training strategies and architectures to a dataset.The related work notes that adequately designed pipelines can make the basic encoder-decoder structure difficult to surpass.Source: §2.1 · p. 4
2.2. Semi-supervised Medical Image Segmentation
- Semi-supervised methods use pseudo-labels or consistency learning; pseudo-label quality can be poor, while task-level regularization can incorporate geometric information through distance maps.Prior examples include dual-task consistency and mutual learning of dual-task networks.Source: §2.2 · pp. 4–5
2.3. Uncertainty Estimation
- Uncertainty estimates can help identify confident predictions and guide semi-supervised models toward relatively reliable information from unlabeled data.The cited approaches derive uncertainty from ensembles or Monte Carlo dropout, among other methods.Source: §2.3 · p. 5
3. Methodology
The methodology combines labeled and unlabeled cases in a dual-task segmentation framework. It uses uncertainty-guided intra-task consistency and cross-task regularization alongside supervised learning.
3. Methodology
- The framework uses labeled and unlabeled cases to train a data-efficient model, aiming for performance comparable to a model trained on fully labeled data.The labeled set pairs images with ground truth; unlabeled images contribute without corresponding labels.Source: §3 · p. 5
3.1. Backbone Network Architecture and Supervised Learning
- A segmentation branch predicts probabilistic maps while an auxiliary regression branch predicts signed distance maps, providing two task perspectives on target geometry.Signed distance values are negative inside the target, positive outside, and have magnitudes set by distance to the closest boundary.Source: §3.1 · p. 6
3.2. Intra-Task Consistency Regularization
- The mean-teacher framework applies consistency learning to both segmentation and regression outputs, using estimated uncertainty to select relatively certain predictions.Uncertainty is estimated with Monte Carlo Dropout across stochastic teacher-model passes; the teacher weights are an exponential moving average of student weights.Source: §3.2 · pp. 6–7
3.3. Cross-Task Consistency Regularization
- Cross-task regularization aligns segmentation predictions with regression outputs mapped back into segmentation space, exploiting consistency between the two task representations.The method uses a smooth approximation to the inverse signed-distance transform for this mapping.Source: §3.3 · pp. 7–8
3.4. Overall Training Procedure
- The training objective combines supervised segmentation, intra-task consistency, and cross-task consistency losses, with ramp-up weights balancing supervised and unsupervised losses.The ramp-up weights are intended to mitigate consistency-loss disturbance early in training.Source: §3.4 · pp. 8–9
4. Experiments
Experiments evaluated the framework on LA and BraTS with labeled and unlabeled scans, using V-Net and ablations of its consistency components. The reported comparisons favor the proposed method over supervised and semi-supervised alternatives on both datasets.
4.1. Datasets and Experimental Setup
- The LA experiments used 16 labeled and 64 unlabeled scans from an 80-scan training split; BraTS used 25 or 50 labeled scans from 250 training scans.The LA dataset had 100 scans split into 80 training and 20 testing; BraTS had 335 patients, with 60 scans reserved for testing.Source: §4.1 · p. 9
4.2. Implementing Details and Evaluation Metrics
- Experiments used V-Net, SGD, 6,000 training iterations, and four metrics: Dice, Jaccard, average surface distance, and 95% Hausdorff distance.The same V-Net backbone was used across experiments for fair comparison.Source: §4.2 · pp. 9–10
4.3. Ablation Experiments
- The consistency balance weight β performed best at 0.75; using only segmentation or regression consistency produced lower performance.The experiments varied β to balance the two tasks in self-ensembling.Source: §4.3 · p. 10, Table 2 · p. 10
- For the uncertainty threshold, the framework used a ramp from 3/4 Umax to Umax rather than fixing it at Umax throughout training.The threshold choices were compared using Dice on the LA dataset.Source: Figure 3 · p. 11, §4.3 · p. 12
- On LA, distance-map supervision improved Dice by 0.47% and Jaccard by 0.68%, while integrating intra-task and cross-task consistency further promoted performance.The ablation improvements were reported as statistically significant at p < 0.05.Source: Table 1 · p. 10, §4.3 · pp. 12–13
4.4. Performance of Using Different Percentages of Labeled Data
- With 20% labeled LA data, the method improved Dice by 4.13% over the supervised baseline, and its advantage narrowed as the labeled-data share increased.The comparison covered labeled-data shares from 5% to 50%.Source: Figure 4 · p. 12, §4.4 · p. 13
4.5. Comparison Experiments with Other Semi-Supervised Segmentation Methods on LA Dataset
- On LA, Dice rose from 86.03% to 90.16% and Jaccard from 76.06% to 82.18%; the method also outperformed other semi-supervised methods on all four metrics.Its 90.16% Dice was comparable to the 91.14% fully supervised upper-bound result.Source: Table 3 · p. 11, §4.5 · p. 14
4.6. Comparison Experiments with Other Semi-Supervised Segmentation Methods on BraTS Dataset
- On BraTS, the method outperformed other state-of-the-art semi-supervised methods with both 10% and 20% labeled data, while their performances were described as comparable.The supervised baseline averaged 73.00% Dice with 10% labeled data and 76.14% with 20%.Source: Table 4 · p. 14, §4.6 · p. 15
5. Conclusion
The framework combines intra-task and cross-task consistency learning, using estimated uncertainty to select relatively certain predictions from unlabeled data. On two public datasets, it improves on supervised baselines and outperforms other semi-supervised methods under the same backbone and task settings, while its evaluation is limited to single-class tasks on relatively small datasets.
5. Conclusion
- The framework combines intra-task consistency from up-to-date predictions with cross-task regularization for geometric shape information, selecting relatively certain predictions using estimated segmentation uncertainty.It uses a dual-task backbone to produce segmentation probabilistic maps and signed distance maps simultaneously.Source: §5 · p. 15
- Up to 4.13% and 9.82% in Dice coefficient over the supervised baseline on left atrium and brain tumor segmentation, respectively.The reported gains come from leveraging unlabeled data.Source: §5 · p. 16
- The method achieves better segmentation performance than other semi-supervised methods under the same backbone and task settings on both datasets.The conclusion reports experiments on two public medical image segmentation datasets.Source: §5 · p. 16
- The evaluation is limited to single-class segmentation tasks on relatively small datasets; future work targets more challenging multi-class tasks and more diverse datasets.The stated aim is to further validate performance for real-world clinical applications.Source: §5 · p. 16