Information bottleneck-based interpretable multitask network for breast cancer classification and segmentation.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Breast cancer is a leading cause of death in women globally.
• Accurate interpretation of breast imaging requires significant expertise.
• Current deep learning models for breast cancer lack interpretability and often neglect task correlations.

Purpose of the Study:

• To develop an interpretable deep learning framework for simultaneous breast tumor classification and segmentation.
• To address the 'black box' nature of existing algorithms by providing decision-making transparency.
• To enhance the correlation between different diagnostic tasks in breast imaging analysis.

Main Methods:

• Proposed an interpretable multitask information bottleneck network (MIB-Net).
• MIB-Net maximizes mutual information between latent representations and class labels while minimizing information from inputs.
• Implemented multitask learning with a dual prior knowledge guidance strategy for enhanced task correlation.

Main Results:

• MIB-Net generates contribution score maps for interpretable decision-making.
• The framework demonstrated improved accuracy in breast tumor classification and segmentation compared to state-of-the-art models.
• Evaluations were conducted on three diverse breast image datasets.

Conclusions:

• The proposed MIB-Net offers an interpretable solution for breast cancer diagnosis.
• The model enhances physician understanding of AI-driven predictions.
• MIB-Net achieves superior performance in multitask breast imaging analysis.