Towards Low-Cost Wireless AI-assisted Breast Tumor Screening and Volumetric Freehand Reconstruction

Breast ultrasound is valuable for patients with dense tissue and suspicious findings, but conventional systems remain limited by operator dependence, clinic-based infrastructure, and cost. These constraints restrict scalability for frequent imaging and deployment outside clinical settings. Advances in wireless handheld technology create opportunities for patient-performed imaging, but automated interpretation under constrained acquisition conditions remains unexplored. We present a low-cost, wireless, AI-assisted framework for self-directed breast ultrasound using a consumer-grade handheld probe with a mobile tablet interface. Rather than clinical screening performance, this work evaluates whether deep learning models trained on clinical data can generate meaningful predictions on patient-acquired wireless scans. Wireless imaging introduces technical challenges, including reduced channel count, simplified beamforming, lower signal-to-noise ratio, and acquisition variability, resulting in pronounced domain shift from conventional clinical data. We evaluated ResNet50, ResNeXt, and VGG16 for frame-level classification using over 5,000 frames from public datasets, institutional imaging, and wireless recordings, labeled as suspicious or non-suspicious. ResNeXt achieved the most stable performance (88.3% accuracy, 0.9692 AUC-PR). Applied to patient-acquired wireless cine loops from predefined freehand trajectories, model outputs produced localized suspicious prediction clusters showing qualitative spatial agreement with documented lesion locations. We also present preliminary IMU-based pose estimation toward volumetric reconstruction. This work demonstrates that AI-assisted interpretation can enable scalable, low-cost ultrasound workflows and informs engineering development of patient-accessible imaging systems.