Multi-scale Information Aggregation Network for Spine MRI Image Segmentation∗

Intervertebral disc herniation, spinal stenosis, and degenerative disc are spinal diseases with a high incidence. Accurate segmentation of spinal images is crucial for the diagnosis and treatment of related diseases. This paper proposes a multi-scale information aggregation U-shaped network (MIAU-Net) for spinal magnetic resonance images. MIAU-Net is a novel semantic segmentation model which improved on the U-Net. This model gets better segmentation performance by redesigning the encoder-decoder and the skip connection module. Specifically, the proposed multi-scale information aggregation module is used to capture features of different scales through different receptive fields. While the redesigned skip connection module can speed up the training process and alleviate the problem of gradient disappearance. The model is evaluated using the publicly available SpineSagT2Wdataset3 spine image dataset. Evaluation metrics include the Dice similarity coefficient (DSC), intersection over union, true positive rate, positive predictive value, and F1 score. The DSC score can reach 90.41%. Comparing with other state-of-the-art networks can verify that this method realizes more accurate semantic segmentation of the spine.