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

Breast cancer remains a leading cause of cancer-related mortality among women, and breast ultrasound (BUS) offers a widely used, non-invasive modality for early detection and treatment planning. However, BUS-based computer-aided diagnosis is often hindered by speckle noise, device-dependent intensity variations, and substantial heterogeneity in lesion size, contrast, and appearance, which can degrade the robustness of conventional CNN classifiers. In this study, we propose a multi-scale information-bottleneck-guided classification framework tailored to BUS lesion analysis. A ResNet backbone with a feature pyramid network (FPN) extracts hierarchical multi-scale representations that jointly encode fine lesion details and global structural context. An information bottleneck (IB) module is attached to each FPN level to learn channel-wise gating masks from global statistics, suppressing background glandular textures and scanner-related artifacts while preserving discriminative lesion cues; auxiliary classifiers provide deep supervision at each scale, producing per-level class-posterior estimates. These scale-specific predictions are then aggregated via confidence-weighted decision-level fusion to obtain a unified, uncertainty-aware output. Experiments on breast ultrasound datasets demonstrate improved classification performance and enhanced robustness over representative CNN baselines, with notable benefits in challenging small and low-contrast lesion scenarios, supporting the framework's potential for clinically applicable BUS-based screening.