RS²-Net: An end-to-end deep learning framework for rodent skull stripping in multi-center brain MRI

Skull stripping is a crucial preprocessing step in magnetic resonance imaging (MRI), where experts manually create brain masks. This labor-intensive process heavily relies on the annotator's expertise, as automation faces challenges such as low tissue contrast, significant variations in image resolution, and blurred boundaries between the brain and surrounding tissues, particularly in rodents. In this study, we have developed a lightweight framework based on Swin-UNETR to automate the skull stripping process in MRI scans of mice and rats. The primary objective of this framework is to eliminate the need for preprocessing, reduce the workload, and provide an out-of-the-box solution capable of adapting to various MRI image resolutions. By employing a lightweight neural network, we aim to lower the performance requirements of the framework. To validate the effectiveness of our approach, we trained and evaluated the network using publicly available multi-center data, encompassing 1,037 rodents and 1,142 images from 89 centers, resulting in a preliminary mean Dice coefficient of 0.9914. The framework, data, and pre-trained models can be found on the following link: https://github.com/VitoLin21/Rodent-Skull-Stripping.