ScribbleVS: Scribble-Supervised Medical Image Segmentation via Dynamic Competitive Pseudo Label Selection

TL;DR

ScribbleVS tackles the high annotation cost of medical image segmentation by exploiting scribble annotations and addressing pseudo-label noise. It introduces Regional Pseudo Label Diffusion (RPD) to propagate reliable supervision from sparse scribbles to uncertain regions and Dynamic Competitive Selection (DCS) to choose the most trustworthy predictions during training within a mean-teacher framework. The method jointly optimizes a scribble-based loss and a regionally restricted pseudo-label loss, with a Gaussian warm-up for balancing supervision signals. Across four public datasets (ACDC, MSCMRseg, WORD, BraTS2020), ScribbleVS achieves segmentation performance close to or exceeding fully supervised baselines and outperforms prior scribble-based and weakly supervised methods, demonstrating robust generalization and labeling efficiency.

Abstract

In clinical medicine, precise image segmentation can provide substantial support to clinicians. However, obtaining high-quality segmentation typically demands extensive pixel-level annotations, which are labor-intensive and expensive. Scribble annotations offer a more cost-effective alternative by improving labeling efficiency. Nonetheless, using such sparse supervision for training reliable medical image segmentation models remains a significant challenge. Some studies employ pseudo-labeling to enhance supervision, but these methods are susceptible to noise interference. To address these challenges, we introduce ScribbleVS, a framework designed to learn from scribble annotations. We introduce a Regional Pseudo Labels Diffusion Module to expand the scope of supervision and reduce the impact of noise present in pseudo labels. Additionally, we introduce a Dynamic Competitive Selection module for enhanced refinement in selecting pseudo labels. Experiments conducted on the ACDC, MSCMRseg, WORD, and BraTS2020 datasets demonstrate promising results, achieving segmentation precision comparable to fully supervised models. The codes of this study are available at https://github.com/ortonwang/ScribbleVS.

Figures (10)

• Figure 1: Examples of pixel level annotation and scribble annotations. BG, RV, Myo, LV, and UA represent the background, right ventricle, myocardium, left ventricle, and unannotated pixels respectively.
• Figure 2: The schematic diagram of the proposed ScribbleVS framework, which comprises the $RPD$ and the $DCS$ module. The $\mathcal{PL}$ denotes the pseudo labels, and the areas in the $\mathcal{PL}$ indicate regions excluded from loss computation. EMA refers to the exponential moving average mechanism, through which the teacher network is updated based on the student network’s parameters.
• Figure 3: Visualization with other methods on the ACDC and MSCMRseg datasets. The models are trained with scribble annotations.
• Figure 4: Visual comparison between our method and the other weakly supervised methods on the WORD dataset. From top to botton, the visualizations correspond to axial, coronal, sagittal, and 3D views.
• Figure 5: Visual comparison between our method and the other weakly supervised methods on the BraTS2020 dataset. From top to botton, the visualizations correspond to axial, coronal, sagittal, and 3D views.