A dual-path neural network fusing dual-sequence magnetic resonance image features for detection of placenta accrete spectrum (PAS) disorder

With the increase of various risk factors such as cesarean section and abortion, placenta accrete spectrum (PAS) disorder is happening more frequently year by year. Therefore, prenatal prediction of PAS is of crucial practical significance. Magnetic resonance imaging (MRI) quality will not be affected by fetal position, maternal size, amniotic fluid volume, etc., which has gradually become an important means for prenatal diagnosis of PAS. In clinical practice, T2-weighted imaging (T2WI) magnetic resonance (MR) images are used to reflect the placental signal and T1-weighted imaging (T1WI) MR images are used to reflect bleeding, both plays a key role in the diagnosis of PAS. However, it is difficult for traditional MR image analysis methods to extract multi-sequence MR image features simultaneously and assign corresponding weights to predict PAS according to their importance. To address this problem, we propose a dual-path neural network fused with a multi-head attention module to detect PAS. The model first uses a dual-path neural network to extract T2WI and T1WI MR image features separately, and then combines these features. The multi-head attention module learns multiple different attention weights to focus on different aspects of the placental image to generate highly discriminative final features. The experimental results on the dataset we constructed demonstrate a superior performance of the proposed method over state-of-the-art techniques in prenatal diagnosis of PAS. Specifically, the model we trained achieves 88.6% accuracy and 89.9% F1-score on the independent validation set, which shows a clear advantage over methods that only use a single sequence of MR images.