Automatic brain extraction for rat magnetic resonance imaging data using U²-Net

Abstract

Objective. Skull stripping is a key step in the pre-processing of rodent brain magnetic resonance images (MRI). This study aimed to develop a new skull stripping method via U2-Net, a neural network model based on deep learning method, for rat brain magnetic resonance images.
Approach. In this study, 599 rats were enrolled and U2-Net was applied to segment MRI images of rat brain. The intercranial tissue of each rat was manually labeled. 476 rats (approximate 80%) were used for training set while 123 rats (approximate 20%) were used to test the performance of the trained U2-Net model. For evaluation, the segmentation result by the U2-Net model is compared with the manual label, and traditional segment methods. Quantitative evaluation, including Dice coefficient, Jaccard coefficient, Sensitivity, Specificity, Pixel accuracy, Hausdorff coefficient, True positive rate, False positive rate and the volumes of whole brain, were calculated to compare the segmentation results among different models.
Main results. The U2-Net model was performed better than the software of RATS and BrainSuite, in which the quantitative values of training U2-Net model were 0.9907±0.0016(Dice coefficient), 0.9816±0.0032(Jaccard coefficient), 0.9912±0.0020(Sensitivity), 0.9989±0.0002(Specificity), 0.9982±0.0003( Pixel accuracy), 5.2390±2.5334(Hausdorff coefficient), 0.9902±0.0025(True positive rate), 0.0009±0.0002(False positive rate) respectively. 
Significance. This study provides a new method that achieves reliable performance in rat brain skull stripping of MRI images, which could contribute to the processing of rat brain magnetic resonance images.