Multi-stage deep learning model for the classification and segmentation of breast ultrasound sweeps by non-experts

Breast cancer remains a leading cause of mortality among women globally, with under-resourced regions facing significant barriers to early diagnosis. To address this, we propose an automated, radiologist-free system for analyzing palpable breast lumps using low-cost ultrasound and volume sweep imaging. We introduce a novel frame-based, multi-stage deep learning pipeline that performs both segmentation and classification of breast tumors in ultrasound cine clips. Our approach first isolates tumor regions using wavelet attention UNet, then fuses the original image, tumor-isolated region, and mask into a composite RGB input to enrich spatial context. A bagging ensemble of ResNet152v2, EfficientNetV2-L, and DenseNet201 performs classification on these enhanced inputs. The model was trained and tested on two datasets: 3818 frames from a portable Butterfly iQ probe (VSI dataset) and 780 images from the public breast ultrasound image dataset. The RGB fusion approach led to a 10% improvement in accuracy and substantial gains in sensitivity and specificity compared to grayscale baselines. Performance gains were statistically significant (adjusted $p \lt 1\times 10^{-4}$ , McNemar’s test). A clinical-scale evaluation against pathology-confirmed diagnoses demonstrated 95% sensitivity in malignant cases. When applied to full sweep summaries, the model achieved 73% accuracy, highlighting its real-world utility on video data. This study presents the first multi-stage segmentation-classification framework for portable breast ultrasound cine sweeps, validated against clinical ground truth. The system offers a promising, low-cost screening solution for breast cancer detection in low- and middle-income countries.