Scribble-Supervised Multi-Organ Segmentation via Epistemic-Driven Hardness-Adaptive Focusing.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Scribble supervision in multi-organ segmentation reduces annotation costs but suffers from sparsity, leading to poor feature learning in difficult areas like organ boundaries.
• This sparsity causes model confirmation bias and high epistemic uncertainty, which current methods do not adequately address.

Purpose of the Study:

• To propose an epistemic-driven hardness-adaptive focusing framework to overcome limitations of scribble supervision in multi-organ segmentation.
• To reduce model confirmation bias and epistemic uncertainty in hard-to-segment regions.

Main Methods:

• Developed a phase-adaptive hardness-aware loss function to quantify epistemic uncertainty and generate dynamic hardness maps.
• Employed a distribution-divergence-aware copy-paste operation for hard sample generation and progressive learning.
• Introduced feature distribution alignment to mitigate bias and uncertainty by aligning organ-specific hard regions with global features.

Main Results:

• The proposed framework demonstrated competitive performance and effectiveness on multi-organ CT and ultrasound datasets.
• Generalizability and robustness were validated across cross-dataset and noise-corrupted scenarios.
• The method offers a practical solution for efficient annotation in clinical applications.

Conclusions:

• The epistemic-driven hardness-adaptive focusing framework effectively improves multi-organ segmentation accuracy with limited annotations.
• The self-improving loop of uncertainty quantification, hard sample generation, and feature alignment successfully reduces bias and epistemic uncertainty.
• This approach provides a valuable tool for clinical settings prioritizing annotation efficiency.