Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PLESS: Pseudo-Label Enhancement with Spreading Scribbles for Weakly Supervised Segmentation

Yeva Gabrielyan, Varduhi Yeghiazaryan, Irina Voiculescu

TL;DR

PLESS tackles the challenge of noisy and incomplete supervision in scribble-based weakly supervised segmentation by introducing a hierarchical, region-based scribble propagation strategy. By partitioning images into semantically coherent regions via watershed and waterfall transforms, propagating scribble information within regions, and adding a background expansion step, PLESS refines pseudo-labels used in training. The method is model-agnostic and can be integrated with existing pseudo-label losses; experiments on two cardiac MRI datasets show consistent improvements across four scribble-based baselines, with notable gains in boundary accuracy. The work highlights the value of encoding spatial coherence in pseudo-label refinement and suggests broad applicability to other sparse/noisy annotation settings beyond medical imaging.

Abstract

Weakly supervised learning with scribble annotations uses sparse user-drawn strokes to indicate segmentation labels on a small subset of pixels. This annotation reduces the cost of dense pixel-wise labeling, but suffers inherently from noisy and incomplete supervision. Recent scribble-based approaches in medical image segmentation address this limitation using pseudo-label-based training; however, the quality of the pseudo-labels remains a key performance limit. We propose PLESS, a generic pseudo-label enhancement strategy which improves reliability and spatial consistency. It builds on a hierarchical partitioning of the image into a hierarchy of spatially coherent regions. PLESS propagates scribble information to refine pseudo-labels within semantically coherent regions. The framework is model-agnostic and easily integrates into existing pseudo-label methods. Experiments on two public cardiac MRI datasets (ACDC and MSCMRseg) across four scribble-supervised algorithms show consistent improvements in segmentation accuracy. Code will be made available on GitHub upon acceptance.

PLESS: Pseudo-Label Enhancement with Spreading Scribbles for Weakly Supervised Segmentation

TL;DR

PLESS tackles the challenge of noisy and incomplete supervision in scribble-based weakly supervised segmentation by introducing a hierarchical, region-based scribble propagation strategy. By partitioning images into semantically coherent regions via watershed and waterfall transforms, propagating scribble information within regions, and adding a background expansion step, PLESS refines pseudo-labels used in training. The method is model-agnostic and can be integrated with existing pseudo-label losses; experiments on two cardiac MRI datasets show consistent improvements across four scribble-based baselines, with notable gains in boundary accuracy. The work highlights the value of encoding spatial coherence in pseudo-label refinement and suggests broad applicability to other sparse/noisy annotation settings beyond medical imaging.

Abstract

Weakly supervised learning with scribble annotations uses sparse user-drawn strokes to indicate segmentation labels on a small subset of pixels. This annotation reduces the cost of dense pixel-wise labeling, but suffers inherently from noisy and incomplete supervision. Recent scribble-based approaches in medical image segmentation address this limitation using pseudo-label-based training; however, the quality of the pseudo-labels remains a key performance limit. We propose PLESS, a generic pseudo-label enhancement strategy which improves reliability and spatial consistency. It builds on a hierarchical partitioning of the image into a hierarchy of spatially coherent regions. PLESS propagates scribble information to refine pseudo-labels within semantically coherent regions. The framework is model-agnostic and easily integrates into existing pseudo-label methods. Experiments on two public cardiac MRI datasets (ACDC and MSCMRseg) across four scribble-supervised algorithms show consistent improvements in segmentation accuracy. Code will be made available on GitHub upon acceptance.
Paper Structure (6 sections, 2 equations, 3 figures, 4 tables)

This paper contains 6 sections, 2 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Pseudo-label Enhancement
  3. Experimental Setup
  4. Results
  5. Ablation Studies
  6. Conclusion

Figures (3)

  • Figure 1: Overview of the proposed PLESS framework, showing pseudo-label construction, waterfall-based enhancement, and integration into pseudo-label loss. The red dotted arrows indicate the original pseudo-label DSC loss formulation used prior to PLESS, which is replaced by the green-arrow loss terms introduced by the proposed enhancement strategy.
  • Figure 2: Comparison of segmentation performance using the baseline and PLESS on three scans (a, b, c) from the ACDC dataset. Row 1 displays the original MRI slice and ground truth (GT), while subsequent rows show segmentation results from different models: DMPLS luo2022scribble, DCDPL wang2023weakly, ScribbleVC li2023scribblevc, and ScribbleVS wang2024scribblevs. Reported DSC, HD95, and ASD scores refer to the whole 3D scan. Label colors: background , right ventricle (RV) , myocardium (MYO) , and left ventricle (LV) .
  • Figure 3: Qualitative comparison among enhancement results with different PLESS setups on a sample from ACDC.