Category-specific unlabeled data risk minimization for ultrasound semi-supervised segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate ultrasound image analysis is hindered by artifacts and the scarcity of expert annotations.
• Semi-supervised segmentation offers a solution but often overlooks tissue-specific learning challenges and empirical risk minimization for unlabeled data.

Purpose of the Study:

• To propose a novel semi-supervised segmentation model, CSUDRM, addressing limitations in current ultrasound image analysis.
• To enhance model generalization and robustness by incorporating category-specific distribution alignment and unlabeled data risk minimization.

Main Methods:

• Developed CSUDRM with two modules: Category-Specific Distribution Alignment (CSDA) for consistent feature learning and Unlabeled Data Risk Minimization (UDRM) for optimizing overall training data.
• CSDA enhances intra-class compactness and inter-class discrepancy with category-specific penalties.
• UDRM estimates and minimizes risk on unlabeled data using a learnable class prior estimator guided by CSDA.

Main Results:

• CSUDRM achieved state-of-the-art performance across four diverse ultrasound datasets.
• Ablation studies confirmed the superiority of the proposed CSDA and UDRM modules.
• Feature space visualization and robustness analysis demonstrated enhanced model stability and accuracy.

Conclusions:

• The proposed CSUDRM model effectively addresses key challenges in semi-supervised ultrasound image segmentation.
• The integration of CSDA and UDRM significantly improves segmentation accuracy, robustness, and generalization.
• This work provides a robust framework for advancing computer-aided diagnosis in medical imaging.