To propose a multi-scale information-bottleneck-guided classification framework for breast ultrasound lesion analysis that enhances classification performance and addresses challenges such as speckle noise and device-dependent variations.
Approach:
Information Bottleneck Module: IB modules learn channel-wise gating masks to suppress background noise and artifacts while preserving discriminative lesion features.
Key Findings:
The proposed framework improves classification performance over conventional CNN baselines.
Notable enhancements are observed in challenging scenarios involving small and low-contrast lesions.
The method aligns with clinical workflows for breast cancer screening and diagnosis.
Interpretation:
Limitations:
The study does not address the performance of the framework in real-world clinical settings beyond the experimental datasets.
Conclusion:
The proposed multi-scale information-bottleneck-guided classification framework shows potential for improving breast ultrasound lesion classification in challenging cases.
