Lune

CVPR Findings2026

Anatomy-Aware Adaptive Feature Perturbation Framework for Semi-Supervised MRI Segmentation

Ji Lin, Bo Peng, Suping Li, Qianni Zhang

Abstract

Medical image segmentation depends heavily on extensive, precise, and domain knowledge-based annotations, which are costly and labour-intensive. Although semisupervised segmentation can achieve competitive performance with limited annotations, its core consistency regularisation methods typically employ fixed perturbation strategies that apply identical, task-agnostic, and anatomyagnostic perturbations to all pixels or features. Such uniform approaches risk destroying semantic information in anatomically clear regions or failing to accurately depict complex or boundary areas. To address these issues, we propose a hybrid and adaptive feature perturbation framework for semi-supervised MRI segmentation. Specifically, hierarchical hypercolumn features are constructed as anatomical priors to guide spatial feature perturbations. Then, adaptive Gaussian noise is introduced in the first skip connection, where the perturbation intensity is dynamically adjusted according to the similarity of hypercolumn features to preserve key anatomical information and promote discriminative feature learning. In addition, Dropout is applied to the other skip connections to enhance perturbation diversity by adding channel-level feature perturbation. Extensive experiments on the ACDC and Promise 12 datasets demonstrate that our method achieves great performance across nearly all evaluation metrics, offering an effective solution for medical image segmentation with limited annotations.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

Questions to start from

Lune can draw on

the full paper · related work

Get the answer

Free to start. No credit card required.

Lineage

Builds onCited byLeft to right by year; the more cited, the higher.

Builds on12

  1. Supervised Contrastive LearningPrannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna et al.NeurIPS 2020 · 7,049 citations
  2. A ConvNet for the 2020sZhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer et al.CVPR 2022 · 6,782 citations
  3. Perturbed and Strict Mean Teachers for Semi-supervised Semantic SegmentationYuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu et al.CVPR 2022 · 287 citations
  4. Dual Student: Breaking the Limits of the Teacher in Semi-Supervised LearningZhanghan Ke, Daoye Wang, Qiong Yan, Jimmy S. J. Ren et al.ICCV 2019 · 259 citations
  5. Rethinking Semi-Supervised Medical Image Segmentation: A Variance-Reduction PerspectiveChenyu You, Weicheng Dai, Yifei Min, Fenglin Liu et al.NeurIPS 2023 · 147 citations

Related papers

  1. Bidirectional Channel-selective Semantic Interaction for Semi-Supervised Medical SegmentationKaiwen Huang, Yizhe Zhang, Yi Zhou, Tianyang Xu et al.AAAI 2026
  2. GapMatch: Bridging Instance and Model Perturbations for Enhanced Semi-Supervised Medical Image SegmentationWei Huang, Lei Zhang, Zizhou Wang, Yan WangAAAI 2025 · 8 citations
  3. VQ-Seg: Vector-Quantized Token Perturbation for Semi-Supervised Medical Image SegmentationSicheng Yang, Zhaohu Xing, Lei ZhuNeurIPS 2025 · 3 citations
  4. Semi-supervised Medical Image Segmentation through Dual-task ConsistencyXiangde Luo, Jieneng Chen, Tao Song, Guotai WangAAAI 2021 · 754 citations
  5. Adaptive Bidirectional Displacement for Semi-Supervised Medical Image SegmentationHanyang Chi, Jian Pang, Bingfeng Zhang, Weifeng LiuCVPR 2024
Dusk over the sea between two cliffs drawn in fine vertical lines