Brain Extraction from Magnetic Resonance Images Using UNet modified with Residual and Dense Layers

Abstract

The main goal of brain extraction is to separate the brain from non-brain parts, which enables accurate detection or classification of abnormalities within the brain region. The precise brain extraction process significantly influences the quality of successive neuroimaging analyses. Brain extraction is a challenging task mainly due to the similarity of intensity values between brain and non-brain structure. In this study, a UNet model improved with ResNet50 or DenseNet121 feature extraction layers was proposed for brain extraction from Magnetic Resonance Imaging (MRI) images. Three publicly available datasets (IBSR, NFBS and CC-359) were used for training the deep learning models. The findings of a comparison between different feature extraction layer types added to UNet shows that residual connections taken from ResNet50 is more successful across all datasets. The ResNet50 connections proved effective in enhancing the distinction of weak but significant gradient values in brain boundary regions. In addition, the best results were obtained for CC-359. The improvement achieved with CC-359 can be attributed to its larger number of samples with more slices, indicating that the model learned better. The performance of our proposed model, evaluated using test data, is found to be comparable to the results obtained in the literature.