nnMU-Net：A self-configuring neural network architecture with multi-modality information fusion for medical MR image segmentation

Motivation: Leveraging multi-modality information to improve the accuracy of magnetic resonance image (MRI) segmentation plays an important role in clinical application. Goal(s): A general self-configuring neural network structure is developed for accurate medical image segmentation, which could leverage information interaction of multi-modality MRI and automatically configure parameters to accommodate different modalities. Approach: U-Net is adopted as a backbone with self-configuring modifications, and each modality is independently input and then fused together after each encoder/decoder stage to maximize feature utilization. Results: It allows one to automatically configure parameters to accommodate different modalities and improves the segmentation accuracy compared to existing methods. Impact: Our method leverages the features from multi-modality MRI to improve the segmentation accuracy and can automatically configure itself, including preprocessing, network architecture, training and post-processing for different modalities. Also, it provides the application possibility for other multi-modality segmentation tasks.