Source-linked AI summary
Exploring Smoothness and Class-Separation for Semi-supervised Medical Image Segmentation
Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, Jianfei Cai
TL;DR
Semi-supervised medical image segmentation must work with scarce labels and blurred, ambiguous targets. SS-Net combines adversarial perturbation consistency with prototype-based class separation, and experiments on LA and ACDC report superior performance and new state-of-the-art results.
Problem
Semi-supervised medical image segmentation is challenged by scarce labeled data and blurred targets, including ambiguous pixels near adhesive edges and low-contrast regions.
Method
SS-Net jointly enforces pixel-level smoothness under adversarial perturbations and inter-class separation by pulling features toward high-quality class prototypes.
Results
SS-Net outperformed five recent semi-supervised methods and achieved state-of-the-art performance on the LA and ACDC datasets.
Takeaways & Limitations
Combining adversarial smoothness with compact, separated class distributions effectively exploits unlabeled training data for semi-supervised medical image segmentation.
Abstract
from arXiv · showhide
Semi-supervised segmentation remains challenging in medical imaging since the amount of annotated medical data is often scarce and there are many blurred pixels near the adhesive edges or in the low-contrast regions. To address the issues, we advocate to firstly constrain the consistency of pixels with and without strong perturbations to apply a sufficient smoothness constraint and further encourage the class-level separation to exploit the low-entropy regularization for the model training. Particularly, in this paper, we propose the SS-Net for semi-supervised medical image segmentation tasks, via exploring the pixel-level smoothness and inter-class separation at the same time. The pixel-level smoothness forces the model to generate invariant results under adversarial perturbations. Meanwhile, the inter-class separation encourages individual class features should approach their corresponding high-quality prototypes, in order to make each class distribution compact and separate different classes. We evaluated our SS-Net against five recent methods on the public LA and ACDC datasets. Extensive experimental results under two semi-supervised settings demonstrate the superiority of our proposed SS-Net model, achieving new state-of-the-art (SOTA) performance on both datasets. The code is available at https://github.com/ycwu1997/SS-Net.
1 Introduction
Semi-supervised medical segmentation is constrained by scarce pixel labels and ambiguous, blurred targets. SS-Net addresses both challenges by combining adversarial pixel-level smoothness with prototype-based inter-class separation, and reports gains across LA and ACDC evaluations.
- Dense segmentation models require extensive per-pixel annotations, motivating semi-supervised learning with few labeled and abundant unlabeled images.
- Existing semi-supervised methods rely on smoothness under perturbations and low-entropy predictions associated with low-density decision boundaries.
- Medical image segmentation remains difficult because labeled data are limited and targets contain ambiguous pixels near adhesive edges or low-contrast regions.
- SS-Net applies adversarial noises as strong perturbations to enforce pixel-level smoothness and uses labeled high-quality features as class prototypes.
- Prototype-based inter-class separation compacts each class distribution and pushes different classes apart in feature space.
- SS-Net outperforms five recent semi-supervised methods and establishes state-of-the-art performance on the LA and ACDC datasets.
2 Method
SS-Net jointly enforces pixel-level smoothness with adversarial perturbations and inter-class separation with prototype-based feature constraints. Its losses regularize predictions and feature distributions for semi-supervised medical image segmentation.
- Pixel-level Smoothness: SS-Net enforces consistency between original images and adversarially perturbed samples to impose pixel-level smoothness.The adversarial noises provide a stronger smoothness constraint than random noises.
- Pixel-level Smoothness: Dice loss replaces KL divergence for generating adversarial noises and measuring consistency across class predictions.For each class, the loss compares predictions with and without strong perturbations.
- Inter-class Separation: Prototype-based feature constraints address ambiguous targets by shrinking intra-class distributions and separating different classes in feature space.High-quality labeled features form prototypes, while projected features are encouraged to approach their corresponding class prototypes.
- Inter-class Separation: The inter-class separation loss uses cosine similarity to align labeled and unlabeled features with class prototypes.This alignment makes each class distribution compact and improves feature-space separation.
- Overall Objective: The total objective combines labeled-data segmentation loss with smoothness and inter-class separation losses applied to all data.λlds and λcs balance the two regularization terms, while Lseg is a Dice loss for the few labeled samples.
3 Experiment and Results
SS-Net was evaluated on LA and ACDC under 5% and 10% labeled-data settings against five recent methods. Results show strong performance, especially with scarce labels, while ablations support both smoothness and inter-class-separation components.
- Experimental settings: Experiments used fixed LA and ACDC splits with challenging 5% and 10% labeled-data settings, treating remaining training samples as unlabeled.LA used 80 training and 20 validation scans; ACDC used 70 training, 10 validation, and 20 testing scans.
- LA results: SS-Net significantly outperformed other models on LA with 5% labeled training and improved all reported metrics over lower bounds.Evaluation used Dice, Jaccard, 95HD, and ASD, with largest-connected-component post-processing for fair comparison.
- ACDC results: SS-Net achieved the best segmentation performance among compared methods on ACDC with 5% labeled training and reduced most false-positive predictions.ACDC evaluation covered three-class segmentation of the myocardium, left ventricle, and right ventricle.
- ACDC results: With 5% labels, UA-MT reached 46.04% Dice versus 47.83% for the lower bound, indicating a substantial performance decrease in that setting.The authors relate this drop to filtering highly uncertain regions during training.
- Ablation study: Ablations found that either Llds or Lcs improved semi-supervised segmentation, while Dice loss for adversarial-noise estimation achieved 84.31% versus 82.27% Dice with 5% labels.The comparison contrasts Dice loss with K-L Divergence for estimating adversarial noises.
4 Conclusion
The conclusion presents SS-Net as a method that jointly applies adversarial smoothness constraints and class-distribution separation for semi-supervised medical image segmentation. Experiments on LA and ACDC reportedly show superior performance, while future work targets adaptive perturbation and prototype selection.
- Conclusion: SS-Net jointly applies adversarial noises for pixel-level smoothness and shrinks class distributions to separate different classes.The stated motivation is the combination of fewer labels and blurred targets in medical imaging.
- Conclusion: The authors report that this joint strategy can effectively exploit unlabeled training data.This conclusion is stated within the paper’s semi-supervised medical image segmentation setting.
- Conclusion: Experiments on the LA and ACDC datasets reportedly show that SS-Net outperforms other methods and achieves superior semi-supervised segmentation performance.The conclusion summarizes the reported experimental outcome across both datasets.
- Future work: Future work will investigate adaptive selection of perturbation magnitude and prototype size.These are identified as open design choices in the conclusion.
5 Appendix
The appendix specifies procedures for generating adversarial perturbations and constructing class prototypes for the two SS-Net constraints. It also includes the associated cosine-similarity calculation and training outputs.
- Pixel-level smoothness: The pixel-level smoothness algorithm takes a training batch and perturbation magnitude ǫ, then outputs Llds for model training.Its stated input-output interface defines the role of the procedure in the training pipeline.
- Pixel-level smoothness: Adversarial perturbations are generated by initializing random noise, computing Dice-loss gradients, normalizing them, and adding the scaled result to the input.The perturbation is radv ← ǫ · g/∥g∥2 and the adversarial example is xadv = x + radv.
- Inter-class separation: The inter-class-separation algorithm uses labeled and unlabeled deep features, prototype number K, classifier outputs, labels, projectors, and class-specific attention modules.Its output is Lcs for model training.
- Inter-class separation: Labeled features are projected, filtered into correctly classified prototype candidates, ranked by attention, and reduced to the top-K prototype vectors.The procedure then groups prototypes and projected features by true or pseudo labels.
- Inter-class separation: Cosine similarity is used to compute the distance between prototype features and projected features, with class-specific counts N and M appearing in the calculation.The cited appendix text identifies N and M as the numbers of prototype and projected feature vectors in class c.