A Multi-task Approach Using Positional Information for Ultrasound Placenta Segmentation

摘要

Automatic segmentation of the placenta in fetal ultrasound (US) is challenging due to its high variations in shape, position and appearance. Convolutional neural networks (CNN) are the state-of-the-art in medical image segmentation and have already been applied successfully to extract the placenta in US. However, the performance of CNNs depends highly on the availability of large training sets which also need to be representative for new unseen data. In this work, we propose to inform the network about the variability in the data distribution via an auxiliary task to improve performances for under representative training sets. The auxiliary task has two objectives: (ⅰ) enlarging of the training set with easily obtainable labels, and (ⅱ) including more information about the variability of the data in the training process. In particular, we use transfer learning and multi-task learning to incorporate the pla-cental position in a U-Net architecture. We test different models for the segmentation of anterior and posterior placentas in fetal US. Our results suggest that these placenta types represent different distributions. By including the position of the placenta as an auxiliary task, the segmentation accuracy for both anterior and posterior placentas is improved when the specific type of placenta is not included in the training set.