HELPNet: Hierarchical perturbations consistency and entropy-guided ensemble for scribble supervised medical image...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Fully annotating medical images for segmentation is time-consuming and expensive.
• Scribble annotations offer a cost-effective alternative but lack detailed information for accurate organ delineation.

Purpose of the Study:

• To introduce HELPNet, a novel scribble-based weakly supervised segmentation framework.
• To bridge the gap between annotation efficiency and segmentation performance in medical imaging.

Main Methods:

• HELPNet integrates Hierarchical Perturbations Consistency (HPC) for multi-scale feature learning.
• The Entropy-guided Pseudo-Label (EGPL) module generates high-quality pseudo-labels based on prediction confidence.
• Structural Prior Refinement (SPR) module refines pseudo-labels using connectivity and boundary information.

Main Results:

• HELPNet significantly outperforms existing scribble-based weakly supervised segmentation methods.
• Achieved performance comparable to fully supervised segmentation techniques on ACDC, MSCMRseg, and CHAOS datasets.

Conclusions:

• HELPNet offers an efficient and effective solution for medical image segmentation using weak supervision.
• The framework demonstrates the potential of combining hierarchical perturbations, entropy-guided pseudo-labeling, and structural priors for improved segmentation accuracy.