Multi-scale information bottleneck with confidence-weighted decision fusion for robust breast ultrasound lesion classification

1

Breast cancer is a leading cause of cancer-related mortality among women, with breast ultrasound (BUS) being a key non-invasive detection method.
2

Conventional CNN classifiers for BUS face challenges due to speckle noise, intensity variations, and lesion heterogeneity, affecting diagnosis robustness.
3

The proposed framework utilizes a ResNet backbone and feature pyramid network to capture multi-scale representations for improved lesion analysis.
4

Information bottleneck modules are integrated to suppress irrelevant background textures while enhancing lesion-relevant features in BUS images.
5

Experiments demonstrate that the framework outperforms existing CNN baselines, particularly in classifying small and low-contrast breast lesions.

Frontiers In Oncology

1