Answers are grounded in this paper and cite their sources.
Telerobotic ultrasound systems enable clinicians to remotely control robotic arms equipped with ultrasound probes, improving diagnostic access. However, their effectiveness is constrained by limited annotated datasets, privacy concerns associated with centralized storage, and a lack of transparent and generalizable artificial intelligence (AI) models. This study presents a Federated Deep Transfer Learning (FDTL) framework supporting collaborative ultrasound image classification while keeping images on local client nodes. Four pretrained convolutional neural networks (CNNs) were fine-tuned on the breast ultrasound images (BUSI) dataset to extract diagnostic features. Federated averaging (FedAvg) enabled model aggregation without sharing raw client data. Centralized and federated approaches were compared, and convergence was analyzed. A gradient-weighted class activation mapping (Grad-CAM) module supported interpretability. The framework achieved a peak federated classification accuracy of 91.1% using DenseNet-121 and a macro-averaged accuracy of 90.5% across the evaluated architectures, outperforming individual client models while maintaining stable convergence over 80 communication rounds. Convergence stabilized within 50--60 rounds. Grad-CAM highlighted attention regions corresponding to diagnostically relevant lesion boundaries and textures. The framework shows a promising balance of diagnostic performance, data localization, interpretability, and scalability under simulated federated conditions. These findings provide preliminary evidence of robustness across heterogeneous simulated clients and suggest potential for decentralized ultrasound diagnosis.
No takes yet. Share an insight, caveat, or question.
Oluwatobi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: