Source-linked AI summary
Semi-supervised Left Atrium Segmentation with Mutual Consistency Training
Yicheng Wu, Minfeng Xu, Zongyuan Ge, Jianfei Cai, Lei Zhang
TL;DR
Medical image segmentation needs many densely labeled examples, while challenging unlabeled regions may contain important uncertainty information. MC-Net uses two decoders and cycled pseudo labels to encourage mutual consistency, achieving the most accurate reported semi-supervised left atrium segmentation on the LA database while outperforming six recent methods.
Problem
Collecting abundant densely labeled data for medical image segmentation is labor-intensive, and existing methods underestimate challenging unlabeled regions during training.
Method
MC-Net uses one encoder, two slightly different decoders, and a cycled pseudo-label scheme that converts prediction discrepancies into unsupervised mutual-consistency training.
Results
MC-Net outperforms six recent semi-supervised methods and achieves the most accurate semi-supervised left atrium segmentation performance on the LA database.
Takeaways & Limitations
Emphasizing unlabeled challenging regions through consistent, low-entropy predictions supports generalized feature learning for semi-supervised left atrium segmentation.
Takeaways & Limitations
Uncertainty estimation methods such as the cited approach may require multiple inferences, including eight stochastic forward passes per iteration, increasing computational costs.
Abstract
from arXiv · showhide
Semi-supervised learning has attracted great attention in the field of machine learning, especially for medical image segmentation tasks, since it alleviates the heavy burden of collecting abundant densely annotated data for training. However, most of existing methods underestimate the importance of challenging regions (e.g. small branches or blurred edges) during training. We believe that these unlabeled regions may contain more crucial information to minimize the uncertainty prediction for the model and should be emphasized in the training process. Therefore, in this paper, we propose a novel Mutual Consistency Network (MC-Net) for semi-supervised left atrium segmentation from 3D MR images. Particularly, our MC-Net consists of one encoder and two slightly different decoders, and the prediction discrepancies of two decoders are transformed as an unsupervised loss by our designed cycled pseudo label scheme to encourage mutual consistency. Such mutual consistency encourages the two decoders to have consistent and low-entropy predictions and enables the model to gradually capture generalized features from these unlabeled challenging regions. We evaluate our MC-Net on the public Left Atrium (LA) database and it obtains impressive performance gains by exploiting the unlabeled data effectively. Our MC-Net outperforms six recent semi-supervised methods for left atrium segmentation, and sets the new state-of-the-art performance on the LA database.
1 Introduction
Semi-supervised segmentation addresses the labor-intensive need for abundant annotations by exploiting easier-to-collect unlabeled data. MC-Net focuses training on challenging regions using uncertainty-guided mutual consistency and achieves state-of-the-art left atrium segmentation on the LA database.
- Collecting abundant labeled data for medical image segmentation is labor-intensive and time-consuming, motivating effective use of unlabeled data.
- Existing semi-supervised methods use consistency regularization, pseudo labels, teacher-student learning, adversarial alignment, shape regression, and attention mechanisms.
- Ambiguous predictions mainly occur in challenging or blurred regions, which the paper argues should be emphasized because they contain crucial information for training.
- Models trained with more labeled data produce less ambiguous predictions, motivating epistemic uncertainty as an additional supervision signal for unlabeled challenging regions.
- MC-Net uses one encoder, two slightly different decoders, prediction discrepancies, soft pseudo labels, and a cycled pseudo-label scheme to encourage mutual consistency.
- MC-Net explores model-based uncertainty, introduces cycled pseudo labels, and achieves new state-of-the-art performance for semi-supervised left atrium segmentation on the LA database.
2 Method
MC-Net estimates uncertainty through two diverse decoders and converts their prediction discrepancies into mutual consistency supervision. Soft pseudo labels and an unsupervised consistency loss emphasize challenging unlabeled regions while labeled data provide segmentation supervision.
- Model Architecture: MC-Net augments a V-Net-style encoder with two slightly different decoders whose output discrepancy approximates epistemic uncertainty.One decoder uses transposed convolution for up-sampling, while the other uses tri-linear interpolation.
- Model Architecture: Unlike Monte Carlo dropout, MC-Net uses fixed sub-models and avoids additional perturbations during training.This design approximates uncertainty from the discrepancy between decoder outputs PA and PB.
- Cycled Pseudo Label: The method sharpens decoder probabilities into soft pseudo labels that contribute to entropy regularization and reduce the impact of mislabeled training data.The temperature T controls the sharpening function.
- Cycled Pseudo Label: Each decoder supervises the other with its sharpened pseudo labels, encouraging consistent, low-entropy predictions that attend to unlabeled challenging regions.sPLA supervises PB, while sPLB supervises PA, allowing the decoders to learn from each other end-to-end.
- Cycled Pseudo Label: Training minimizes a weighted sum of labeled segmentation loss and unsupervised consistency loss across all training data.The segmentation loss uses labeled data only, whereas the consistency loss supervises all training data; λ balances them.
3 Experiment and Results
Experiments on the LA database evaluate MC-Net under semi-supervised settings and compare its segmentation quality with recent methods and ablated variants. MC-Net performs strongly with limited labeled data and improves results through decoder diversity, consistency, and cycled pseudo-labeling.
- Experimental Setup: The LA database contains 100 gadolinium-enhanced MR scans, split into 80 training and 20 validation samples.The scans have isotropic resolution 0.625 × 0.625 × 0.625 mm.
- Experimental Setup: Training uses randomly cropped 112×112×80 patches, with each batch containing two labeled and two unlabeled patches.Data augmentation includes 2D rotation and flipping.
- Results: MC-Net generates more complete left-atrium segmentations than UA-MT, SASSNet, and DTC under both 10% and 20% labeled-data settings.It also improves challenging areas and removes most isolated regions without morphology-based post-processing.
- Results: 90.34% Dice with 20% labeled training data is comparable to 91.14% for V-Net trained with 100% labeled data.Table 1 reports comparisons using Dice, Jaccard, 95HD, and ASD, and states that MC-Net outperforms six recent semi-supervised methods.
- Ablation Study: Two slightly different decoders outperform two identical decoders, while consistency and cycled pseudo-labeling further improve performance across semi-supervised settings.The ablation study attributes gains to decoder diversity, consistency between pseudo labels, and entropy regularization through the cycled pseudo label.
4 Conclusion
The conclusion presents MC-Net as a mutual consistency approach that emphasizes unlabeled challenging regions in semi-supervised left-atrium segmentation. Its cycled pseudo-label scheme encourages consistent, low-entropy predictions and supports accurate performance on the LA database.
- Conclusion: MC-Net uses a cycled pseudo-label scheme to emphasize unlabeled challenging regions during semi-supervised left-atrium segmentation.The scheme encourages consistent and low-entropy predictions so the model can capture more generalized features from these regions.
- Conclusion: MC-Net achieves the most accurate semi-supervised left-atrium segmentation performance on the LA database, to the authors’ best knowledge.
5 Supplementary Material
Supplementary material compares decoder outputs and extends method comparisons across datasets. The supplied passages describe visual comparisons on LA and table-based comparisons on Pancreas-CT.
- Decoder Comparisons: Table 3 compares the two MC-Net decoder outputs on LA and notes that the original decoder alone achieves superior results without extra computational complexity.
- Cross-Dataset Comparison: Table 4 compares MC-Net with three state-of-the-art methods on the Pancreas-CT database.
- Visual Analysis: Figure 4 displays MR images, both decoder segmentations, their mean MC-Net result, estimated uncertainty, ground truth, and results with or without cycled pseudo labeling.