An unsupervised domain adaptation method for cross-modality spine MRI segmentation

Due to differences in imaging principles and scanning parameters, different modalities of magnetic resonance imaging (MRI) exhibit significant variations in appearance, leading to performance degradation when deploying deep neural networks on unseen modalities. Thus, unsupervised domain adaptation (UDA) between different modalities is an important and challenging task. In this paper, we present a method that combines semantic information guidance and output space feature alignment for unsupervised domain adaptation. It utilizes low-level edge information and foreground region information as guidance to assist the segmentation network in learning domain-invariant features. Subsequently, adversarial learning is employed for cross-domain alignment in the output space, effectively achieving cross-modality segmentation of spine MRI. Testing on public datasets demonstrates that our proposed method outperforms other state-of-the-art methods on all three metrics, exhibiting superior segmentation accuracy.