nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation.

Area of Science:

• Biomedical imaging
• Deep learning
• Medical image analysis

Background:

• Deep learning is advancing biomedical imaging for scientific discovery and medical care.
• Semantic segmentation algorithms are crucial for image analysis but require specialized, dataset-dependent solutions.
• Developing tailored segmentation solutions is complex and resource-intensive.

Purpose of the Study:

• To develop an automated deep learning-based segmentation method for biomedical images.
• To create a tool that requires no manual intervention or specialized expertise.
• To make state-of-the-art image segmentation widely accessible.

Main Methods:

• Developed nnU-Net, a deep learning segmentation framework.
• nnU-Net automatically configures preprocessing, network architecture, training, and post-processing.
• The system uses fixed parameters, interdependent rules, and empirical decisions for self-configuration.

Main Results:

• nnU-Net surpassed existing specialized solutions on 23 public biomedical segmentation datasets.
• The method achieved state-of-the-art performance without manual intervention.
• Demonstrated robust performance across diverse datasets and tasks.

Conclusions:

• nnU-Net provides an out-of-the-box solution for automated biomedical image segmentation.
• The tool democratizes access to advanced deep learning segmentation capabilities.
• Reduces the need for expert knowledge and extensive computational resources.