Speckle aware attention and multi scale pooling for robust breast ultrasound lesion classification across multiple benchmarks

Abstract

Breast ultrasound (BUS) is widely used for lesion assessment, yet automatic benign–malignant classification remains challenging owing to speckle noise, weak boundaries, and cross-scanner heterogeneity. We propose SUA-Net, a compact framework that augments an EfficientNet-B0 backbone with three purpose-built modules: Dual-Path Lightweight Aggregation (DLA) for multi-scale contextual refinement, Speckle Variance Attention (SVA) for variance-guided feature reweighting, and a Multi-Granularity Pooling (MGP) head for enriched global summarization. Predictive uncertainty is estimated via Monte Carlo Dropout at inference time. The full model contains 12.225M parameters and requires 0.630G MACs and 1.260G FLOPs for a 224 × 224 input, with a deterministic model-only latency of 9.708 ± 0.508 ms on an NVIDIA Tesla T4. Under five-fold cross-validation on BUSI, BUSBRA, and BUS-UCLM (image-wise for BUSI and patient-wise for BUSBRA and BUS-UCLM), SUA-Net achieved accuracy/AUC-ROC of 0.940/0.971 on BUSI, 0.870/0.909 on BUSBRA, and 0.830/0.852 on BUS-UCLM. Cross-dataset evaluation showed that BUSBRA-trained models generalised most effectively (AUC-ROC of 0.891 on BUSI and 0.899 on BUS-UCLM), and ablation analysis showed dataset-dependent contributions from DLA, SVA, and MGP. The results support variance-guided attention, lightweight aggregation, and enriched pooling within a compact BUS classification framework and support evaluation across multiple datasets.