A lightweight federated deep learning framework for privacy-preserving fetal ultrasound analysis

The Sustainable Development Goals (SDGs) are concerned with the prevention of stillbirths and promotion of maternal issues to achieve Good Health and Well-being. Ultrasound technology is a crucial component in maternal services because it is a safe, non-invasive, and real-time method of observing fetal status in respect to possible abnormalities. There have been improvements in deep learning, which has enabled efficient analysis of medical images. Current models are complex, which is a barrier to deploying them in Ultrasound devices in limited resource environments. In this study, we propose the Fetal MoAt Net for addressing a multi-class diagnostic plane classification task. The task involves identifying which anatomical diagnostic plane is captured in a given 2D fetal ultrasound image. Fetal MoAt Net is a deep learning model consisting of the MobileNetV2 structure along with the proposed MoAt module to focus on the diagnostically important areas and suppress the ultrasound image-specific noise. The centralized Fetal MoAt Net achieves accuracy of 96.0% (AUC: 99.2%) on the FPUS23 dataset and 91.9% (AUC: 99.3%) on the FETAL_PLANES_DB dataset. The federated model achieves 94.7% (AUC: 98.6%) on FPUS23 and 89.9% (AUC: 97.8%) on FETAL_PLANES_DB, while preserving patient data privacy. The proposed Fetal MoAt Net outperforms the conventional deep learning models such as ResNet-101, DenseNet-121, VGG-16, InceptionV3, and Xception with just 4.1 M parameters and 1.2B FLOPs.The Grad-CAM visualization results show consistent attention to meaningful fetal structures. This proposed method strikes an effective balance between accuracy, interpretability, and efficiency for use in mobile, embedded, and at-home fetal surveillance. We also present the federated learning pipeline which facilitates privacy, preserving, multi, institutional training without the need to share the raw ultrasound data and, thus, enhances the robustness of models across diverse clinical settings.