Fetal MoAt Net: a light-weight deep learning model for fetal diagnostic plane classification

The accurate classification of diagnostic planes from fetal ultrasound images into various categories like abdominal circumference, biparietal diameter, femur length, and no plane can reveal the early signs of possible abnormalities during neonatal care. Even with the advancements made by deep learning, specifically in medical imaging, there has been gradual progress in resource-constrained environments due to the limited availability of sophisticated hardware and the lack of adequate computational resources. These factors impose restrictions on the deployment of standard deep-learning models. To overcome this issue, this study introduces Fetal MoAt Net, a light-weight deep learning architecture integrating a MobileNetV2 backbone with a custom attention block designed to enhance diagnostically relevant regions and suppress ultrasound-specific noise. Our model was evaluated on two benchmark datasets: FPUS23 and FETAL_PLANES_DB, achieving an accuracy of 96% and 91.9%, with AUC up to 100% and 99%, respectively. The results demonstrate the high performance of the proposed model with only 4.1 million parameters and 1.2 billion FLOPs, compared to standard dl models like ResNet 101, DenseNet121, VGG16, InceptionV3, and Xception. These characteristics enable applicability in mobile, embedded, and point-of-care ultrasound systems by demonstrating a strong balance between accuracy, interpretability, and efficiency for real-world fetal health monitoring applications.