AcoustiCoRe-Fed: A Drift-Aware, Fair Federated Learning Framework for Point-of-Care Ultrasound Diagnosis

Point-of-care ultrasound (POCUS) models degrade across hospitals because of drift in devices, acquisition protocols, and patient populations. Federated learning (FL) preserves privacy but assumes static distributions, while continual learning (CL) mitigates drift mainly through replay buffers that are impractical for privacy-sensitive, memory-constrained clinics. We propose AcoustiCoRe-Fed, a physics-guided, importance-weighted, replay-free federated learning framework that detects distribution drift using an Acoustic Consistency Residual (ACR) computed from speckle, attenuation, echogenicity, and texture statistics, adapts locally without transmitting patient data through an importance-weighted, replay-free update rule, and aggregates client models using a sovereignty-fair scheme that prevents data-rich hospitals from dominating the global model, with fairness quantified as a variance-based uniformity measure (100 X [1 - coefficient of variation]) of per-hospital diagnostic accuracy. Evaluated on three open-access public breast-ultrasound datasets (BUSI: 780 images across benign, malignant, and normal cases; UDIAT; and BUS-BRA) partitioned into non-IID hospital clients using a strict patient-level split—where all images from a given patient are assigned exclusively to one client and split to eliminate data leakage—AcoustiCoRe-Fed achieves 88% accuracy and an 87% F1-score. This performance exceeds FedAvg (80%/79%) and FedProx (82%/81%) by 6–8 accuracy points while cutting the forgetting measure by 4.5× and 2× relative to FedAvg and replay-based CL, respectively. Furthermore, it reduces communication cost by 37% and inference latency by 19% relative to centralized training, while raising the fairness score from 70 (base model) to 86 after sovereignty-fair aggregation, demonstrating that per-hospital diagnostic accuracy becomes substantially more uniform across institutions. These results show that combining physics-guided drift detection, replay-free adaptation, and fairness-aware aggregation yields a diagnostically robust, communication-efficient, and edge-deployable solution for heterogeneous point-of-care ultrasound diagnosis.