Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Scribble-Supervised Medical Image Segmentation with Dynamic Teacher Switching and Hierarchical Consistency

Thanh-Huy Nguyen, Hoang-Loc Cao, Dat T. Chung, Mai-Anh Vu, Thanh-Minh Nguyen, Minh Le, Phat K. Huynh, Ulas Bagci

TL;DR

SDT-Net addresses scribble-supervised segmentation under sparse annotations by pairing a single student with two teachers and dynamically selecting the more reliable teacher per iteration. The selected teacher provides high-confidence pseudo-labels refined by a pixel-picking step and a Hierarchical Consistency loss aligns the student's multi-level features with the teacher at both low-level and high-level decoder stages. The training optimizes three losses—scribble supervision, pseudo-label supervision, and hierarchical consistency—with EMA updating applied to the teacher weights. On the ACDC and MSCMRseg datasets, SDT-Net achieves state-of-the-art Dice scores (averages around 90.8% and 90.0%), demonstrating robust, anatomically plausible segmentations while reducing annotation burden.

Abstract

Scribble-supervised methods have emerged to mitigate the prohibitive annotation burden in medical image segmentation. However, the inherent sparsity of these annotations introduces significant ambiguity, which results in noisy pseudo-label propagation and hinders the learning of robust anatomical boundaries. To address this challenge, we propose SDT-Net, a novel dual-teacher, single-student framework designed to maximize supervision quality from these weak signals. Our method features a Dynamic Teacher Switching (DTS) module to adaptively select the most reliable teacher. This selected teacher then guides the student via two synergistic mechanisms: high-confidence pseudo-labels, refined by a Pick Reliable Pixels (PRP) mechanism, and multi-level feature alignment, enforced by a Hierarchical Consistency (HiCo) module. Extensive experiments on the ACDC and MSCMRseg datasets demonstrate that SDT-Net achieves state-of-the-art performance, producing more accurate and anatomically plausible segmentation.

Scribble-Supervised Medical Image Segmentation with Dynamic Teacher Switching and Hierarchical Consistency

TL;DR

SDT-Net addresses scribble-supervised segmentation under sparse annotations by pairing a single student with two teachers and dynamically selecting the more reliable teacher per iteration. The selected teacher provides high-confidence pseudo-labels refined by a pixel-picking step and a Hierarchical Consistency loss aligns the student's multi-level features with the teacher at both low-level and high-level decoder stages. The training optimizes three losses—scribble supervision, pseudo-label supervision, and hierarchical consistency—with EMA updating applied to the teacher weights. On the ACDC and MSCMRseg datasets, SDT-Net achieves state-of-the-art Dice scores (averages around 90.8% and 90.0%), demonstrating robust, anatomically plausible segmentations while reducing annotation burden.

Abstract

Scribble-supervised methods have emerged to mitigate the prohibitive annotation burden in medical image segmentation. However, the inherent sparsity of these annotations introduces significant ambiguity, which results in noisy pseudo-label propagation and hinders the learning of robust anatomical boundaries. To address this challenge, we propose SDT-Net, a novel dual-teacher, single-student framework designed to maximize supervision quality from these weak signals. Our method features a Dynamic Teacher Switching (DTS) module to adaptively select the most reliable teacher. This selected teacher then guides the student via two synergistic mechanisms: high-confidence pseudo-labels, refined by a Pick Reliable Pixels (PRP) mechanism, and multi-level feature alignment, enforced by a Hierarchical Consistency (HiCo) module. Extensive experiments on the ACDC and MSCMRseg datasets demonstrate that SDT-Net achieves state-of-the-art performance, producing more accurate and anatomically plausible segmentation.
Paper Structure (11 sections, 10 equations, 2 figures, 3 tables)

This paper contains 11 sections, 10 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. METHODOLOGY
  3. Scribble-supervised learning
  4. Dynamic Teacher Switching
  5. Hierarchical Consistency
  6. Total Loss Function
  7. EXPERIMENTS
  8. Experimental Setup
  9. Experimental Results and Discussion
  10. CONCLUSION
  11. ACKNOWLEDGMENT

Figures (2)

  • Figure 1: Overview of our SDT-Net. The Dynamic Teacher Switching (DTS) module selects a reliable teacher to generate refined pseudo-labels, while Hierarchical Consistency (HiCo) aligns multi-level features between the student and the selected teacher.
  • Figure 2: Qualitative results across different scribble-supervised medical image segmentation methods on the ACDC dataset. Red dashed circles highlight regions of different segmentation