A deep learning pipeline using priori knowledge for automatic evaluation of placenta accreta spectrum disorders with MRI

Placenta accreta spectrum (PAS) is a pathologic condition of placentation associated with significant maternal morbidity and mortality. We enrolled 540 patients from two institutions to build an automatic pipeline for early diagnosis of PAS based on T2W images. An nnU-Net model was trained for automatic segmentation of the placenta, then an image stripe was created, in which utero-placental borderline (UPB) was straightened and centered. The UPB image was fed into a DenseNet-based network together with placental position for PAS diagnosis. The pipeline achieved good performance with AUCs of 0.860 and 0.897 in internal and external test cohorts, respectively.