SPRNet: Automatic Fetal Standard Plane Recognition Network for Ultrasound Images

摘要

Fetal standard plane recognition is a crucial clinical part of prenatal diagnosis. However, it is also a sophisticated, subjective, and highly empirical process. Thus, there is a huge demand for proposing an effective and precise automatic method to help experienced as well as inexperienced doctors to complete this process, efficiently. In order to satisfy this clinical need, we propose an automatic fetal standard plane recognition network called SPRNet. Specifically, we adopt DenseNet as the basic network of SPRNet and implement data-based partial transfer learning on it by weight-sharing strategy. We then train our network with a task dataset (fetal ultrasound images) and a transferring dataset (placenta ultrasound images) so that our network can discover and learn the potential relationship between these two datasets to improve the performance and avoid overfitting. Finally, we achieve automatic fetal standard plane recognition by utilizing the feature extracted from SPRNet. The experimental results indicate that our network can attain an accuracy of 99.00% and perform better than conventional networks.