Mixed attention mechanism multi-task learning for fetal abdominal standard plane recognition and key anatomical structure detection

Abstract In prenatal ultrasound, accurately identifying fetal abdominal ultrasound standard planes (FAUSP) is challenging due to the complexity of anatomical structures. To address this, we developed FAUSP-NET, a multi-task network that integrates a mixed attention mechanism for real-time FAUSP recognition and anatomical structure detection. The network uses a residual backbone for feature extraction, enhanced by an attention mechanism and Large Selective Kernel Block (LSKblock) for better focus on key regions. A Focal_EIoU loss function addresses class imbalance and improves bounding box regression. Trained on 6767 FAUSP images, FAUSP-NET outperforms 24 popular models, achieving [email protected] of 0.961 and [email protected]:0.95 of 0.653 in detection, with plane recognition accuracy of 0.972. Its average detection time is 24.1 ms. Doctor evaluations show that FAUSP-NET’s accuracy is comparable to senior physicians, offering significant support for clinical ultrasound diagnostics.