Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CAD: Confidence-Aware Adaptive Displacement for Semi-Supervised Medical Image Segmentation

Wenbo Xiao, Zhihao Xu, Guiping Liang, Yangjun Deng, Yi Xiao

TL;DR

This work tackles the challenge of robust semi-supervised medical image segmentation by introducing Confidence-Aware Adaptive Displacement (CAD), which progressively replaces the largest low-confidence regions with high-confidence patches. CAD combines Largest Low-Confidence Region Replacement (LLCR) with Dynamic Threshold Escalation (DTE) to adapt the perturbation scale and confidence threshold throughout training, guided by a Mean Teacher and Cross Pseudo Supervision framework. Empirical results on the ACDC and PROMISE12 datasets demonstrate state-of-the-art performance with limited labeled data, confirming the effectiveness of uncertainty-guided patch replacement. The approach offers a robust, scalable strategy for improving segmentation accuracy and resilience in semi-supervised medical imaging tasks, with plans to release the source code.

Abstract

Semi-supervised medical image segmentation aims to leverage minimal expert annotations, yet remains confronted by challenges in maintaining high-quality consistency learning. Excessive perturbations can degrade alignment and hinder precise decision boundaries, especially in regions with uncertain predictions. In this paper, we introduce Confidence-Aware Adaptive Displacement (CAD), a framework that selectively identifies and replaces the largest low-confidence regions with high-confidence patches. By dynamically adjusting both the maximum allowable replacement size and the confidence threshold throughout training, CAD progressively refines the segmentation quality without overwhelming the learning process. Experimental results on public medical datasets demonstrate that CAD effectively enhances segmentation quality, establishing new state-of-the-art accuracy in this field. The source code will be released after the paper is published.

CAD: Confidence-Aware Adaptive Displacement for Semi-Supervised Medical Image Segmentation

TL;DR

This work tackles the challenge of robust semi-supervised medical image segmentation by introducing Confidence-Aware Adaptive Displacement (CAD), which progressively replaces the largest low-confidence regions with high-confidence patches. CAD combines Largest Low-Confidence Region Replacement (LLCR) with Dynamic Threshold Escalation (DTE) to adapt the perturbation scale and confidence threshold throughout training, guided by a Mean Teacher and Cross Pseudo Supervision framework. Empirical results on the ACDC and PROMISE12 datasets demonstrate state-of-the-art performance with limited labeled data, confirming the effectiveness of uncertainty-guided patch replacement. The approach offers a robust, scalable strategy for improving segmentation accuracy and resilience in semi-supervised medical imaging tasks, with plans to release the source code.

Abstract

Semi-supervised medical image segmentation aims to leverage minimal expert annotations, yet remains confronted by challenges in maintaining high-quality consistency learning. Excessive perturbations can degrade alignment and hinder precise decision boundaries, especially in regions with uncertain predictions. In this paper, we introduce Confidence-Aware Adaptive Displacement (CAD), a framework that selectively identifies and replaces the largest low-confidence regions with high-confidence patches. By dynamically adjusting both the maximum allowable replacement size and the confidence threshold throughout training, CAD progressively refines the segmentation quality without overwhelming the learning process. Experimental results on public medical datasets demonstrate that CAD effectively enhances segmentation quality, establishing new state-of-the-art accuracy in this field. The source code will be released after the paper is published.

Paper Structure

This paper contains 24 sections, 21 equations, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Method
  4. Confidence-Aware Adaptive Displacement
  5. Largest Low-Confidence Region Replacement (LLCR)
  6. Dynamic Threshold Escalation (DTE)
  7. Loss Function
  8. Mean Teacher Loss ($L_\text{MT}$)
  9. Cross Pseudo Supervision Loss ($L_\text{CPS}$)
  10. Confidence-Aware Displacement Loss ($L_\text{CAD}$)
  11. Total Loss
  12. Experiments
  13. Datasets
  14. ACDC Dataset
  15. PROMISE12 Dataset
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Visualization of replacement regions in BCP, ABD, and CAD. Subfigure (a) shows BCP, which swaps foreground and background between labeled and unlabeled data across the entire image. Subfigure (b) illustrates ABD, replacing low-confidence patches with high-confidence patches between strong and weak augmentations. Subfigure (c) demonstrates the proposed CAD, progressively refining replacement regions from small (early stage) to large (late stage) based on confidence levels and training dynamics.
  • Figure 2: Overview of the proposed framework integrating Mean Teacher with DTE, LLCR, and CPS. The diagram illustrates the data flow, confidence-based region replacement, and collaborative learning between the student and teacher models.
  • Figure 3: Segmentation results on three test set samples from the ACDC dataset using 10% labeled data: U-Net, BCP, CorrMatch, ABD, CAD (Ours), and Ground Truth (GT).