Multimodal magnetic resonance imaging (MRI)-based placental implantation detection using deep migration learning for smart biomedical sensing

Background The abnormalities of placental implantation are life-threatening to both the mother and the fetus, and hence the need to detect them accurately and promptly to help in clinical interventions. Multimodal Magnetic Resonance Imaging (MRI) has provided rich anatomical and functional information on the diagnosis of these conditions; however, it is not currently used because of inter-scanner variance and disparity between sets of data. These issues impede the successful application of the MRI-based technologies to detect and assess the risk of Placenta Accreta Spectrum (PAS). Methods To address these limitations, this study proposes a state-of-the-art Multimodal Magnetic Resonance Imaging-based Implantation Classification using a Deep Migration Learning (MIMIC-DML) framework for the automated classification and risk stratification of placenta accreta spectrum (PAS) disorders. The framework integrates multimodal MRI sequences, including T2-weighted, diffusion-weighted (DWI), and contrast-enhanced T1-weighted (T1+C) images, within a deep migration learning architecture for hierarchical feature extraction and cross-domain knowledge transfer. In addition, a domain-adversarial adaptation module is incorporated to enhance cross-scanner robustness, reduce domain discrepancy, and improve model generalization across heterogeneous imaging environments. Results The proposed MIMIC-DML framework achieved superior performance in automated placenta accreta spectrum (PAS) classification, with a classification accuracy of 98.34%, sensitivity of 97.91%, specificity of 98.12%, precision of 97.68%, F1-score of 97.79%, and an AUC of 0.986. In addition, the model demonstrated strong cross-scanner domain adaptability, with stable performance across heterogeneous MRI acquisitions. The segmentation module generated interpretable placental invasion maps with a Dice similarity coefficient of 0.94 and an Intersection over Union (IoU) of 0.89, confirming accurate localization of abnormal implantation regions. These objective findings validate the effectiveness of the framework for robust PAS subtype differentiation and clinically relevant risk stratification. Conclusion The proposed MIMIC-DML framework demonstrates strong potential for accurate placenta accreta spectrum (PAS) detection, subtype-specific risk stratification, and interpretable clinical decision support. By enhancing diagnostic reliability across heterogeneous MRI acquisitions and providing clinically meaningful segmentation outputs, the framework offers significant value for preoperative surgical planning, maternal risk assessment, and early obstetric intervention. These findings establish a robust foundation for future research in domain-adaptive multimodal medical imaging, smart biomedical sensing, and real-time intelligent prenatal diagnostic systems.