Multi-source Medical Sensor Data Fusion Using Transformer for Preoperative Assessment of Placenta Accreta Spectrum

The prenatal diagnosis of placenta accreta spectrum (PAS) is essential to prevent lifethreatening maternal hemorrhage.Recent deep learning models using single-modality ultrasound or magnetic resonance imaging (MRI) have achieved high accuracy, but each modality has inherent limitations: Ultrasound suffers from acoustic shadowing, MRI is prone to motion artifacts, and current methods lack mechanisms to handle missing modalities in routine workflows.To overcome these limitations, we developed a biomimetic medical sensor data fusion model that addresses these limitations through three innovations: (1) a structural prior step using morphological dilation to isolate the placental-myometrial interface, (2) a bidirectional cross-attention module that dynamically aligns acoustic and electromagnetic features, and (3) a missing-modality robust training strategy.On an external validation cohort (N = 60), the model achieved an area under the curve (AUC) of 0.889 [95% confidence interval (CI): 0.824-0.941],outperforming single-modality models (ultrasound's AUC = 0.812; MRI's AUC = 0.843) and previous late-fusion networks (AUC = 0.857).