Boundary sample-based class-weighted semi-supervised learning for malignant tumor classification of medical imaging

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Supervised learning for medical image classification requires extensive labeled data, which is costly and time-consuming to acquire.
• Semi-supervised learning offers an alternative by utilizing limited labeled data with abundant unlabeled data.
• Existing semi-supervised methods struggle with confirmation bias from noisy pseudo-labels and boundary samples.

Purpose of the Study:

• To introduce a novel semi-supervised learning framework, boundary sample-based class-weighted semi-supervised learning (BSCSSL), for medical image classification.
• To mitigate the impact of intra- and inter-class boundary samples in unlabeled medical image data.
• To enhance the accuracy and robustness of medical image classification models.

Main Methods:

• Developed a novel framework named boundary sample-based class-weighted semi-supervised learning (BSCSSL).
• Implemented an inter-class boundary sample mining module to address reliable confidential and inter-class boundary samples.
• Utilized an intra-class boundary sample weighting mechanism to extract class-aware features from difficult-to-classify samples.

Main Results:

• The proposed BSCSSL method demonstrated superior performance compared to existing semi-supervised learning approaches.
• Experimental results on benchmark medical image datasets confirmed the effectiveness of the BSCSSL framework.
• The method successfully alleviated the impact of intra- and inter-class boundary samples.

Conclusions:

• BSCSSL significantly enhances the accuracy and robustness of medical image classification.
• The framework offers valuable contributions to advancing medical diagnosis and related research.
• This approach provides a more efficient way to leverage unlabeled data in medical AI.