A Multi-Orientation Feature Fusion CNN for Gestational Age Prediction from Fetal MRI Images

Fetal magnetic resonance imaging (MRI) sequences are commonly used for identifying fetal growth restriction and disease. In clinical practice, doctors usually rely on the naked eyes to identify and make judgments based on the shape and depth of the sulci and gyri at different gestational ages. Deep learning provides a possible way for automatic gestational age prediction; however, it is difficult to obtain a large training dataset for all gestational weeks. To solve this limitation, this study develops a multi-orientation feature fusion convolutional neural network (MFCNN) for gestational age prediction from fetal MRI images. The multiple rotated instances are generated from the original images, and then input into a Siamese network with shared weight to extract their depth features. The ConvNeXt network is chosen as the backbone architecture of our algorithm. In the experiment, 528 images from 45 patients are used for evaluation, where the gestational ages range from 22 to 34 weeks. The experimental results show that our MFCNN method could improve the performance of baseline ConvNeXt, suggesting that this method has the potential to improve the accuracy of gestational age prediction in the case of a small dataset.