Semi-supervised breast cancer pathology image segmentation based on fine-grained classification guidance

Area of Science:

• Medical image analysis
• Computational pathology
• Artificial intelligence in oncology

Background:

• Breast cancer pathological image segmentation (BCPIS) is crucial for tumor quantification and treatment guidance.
• Fine-grained semantic segmentation is challenging due to complex tissue morphologies and limited, costly manual annotations.
• Existing methods struggle with small, low-quality datasets, impacting segmentation performance.

Purpose of the Study:

• To develop a semi-supervised learning model for accurate BCPIS.
• To address challenges of limited labeled data and complex image features.
• To enhance segmentation performance using classification-guided strategies.

Main Methods:

• Proposed a semi-supervised learning model integrating classification-guided segmentation.
• Utilized a multi-scale convolutional network for feature extraction.
• Employed a multi-expert cross-layer joint learning strategy with pseudo-labeling and data augmentation.

Main Results:

• The semi-supervised model achieved an IoU of 71.53%, demonstrating significant progress.
• The model showed a 3% performance advantage over other semi-supervised methods.
• A strong correlation between classification and segmentation tasks was observed, enhancing segmentation.

Conclusions:

• The proposed semi-supervised model effectively addresses limitations in BCPIS datasets.
• Classification guidance significantly improves fine-grained semi-supervised semantic segmentation.
• This approach offers a promising direction for automated breast cancer image analysis.