Publication

Abstract : Magnetic Resonance Imaging (MRI) is the most extensively used technique for diagnostic assessment of the functioning of the brain. Compared to other neuroimaging techniques like CT and PET, MRI provides more soft tissue variance. Hence MRI provides a proper envision of brain structure. But the presence of tissues which are not part of the brain in the MRI images adversely affect image analysis. Skull stripping removes the non brain tissues in the MRI images of the brain. After skull stripping the analysis of brain region becomes more efficient. Skull stripping can be done in three different ways. Manual skull stripping, by classical segmentation techniques and by deep learning segmentation techniques. The output resulted from manual skull stripping will be used as ground truth for other methods. Lower grade image processing techniques like morphological operations, histogram analysis techniques and region based growing techniques are used in classical segmentation. Skull stripping with deep learning resulted in a better level of accuracy.For performing skull stripping in brain MRI we propose two deep network architectures UNet and ResUNet. UNet network is designed for semantic segmentation. In ResUNet residual learning is added to the UNet. In this study the residual learning is not showing much impact on the performance and UNet shows a better performance over ResUNet . In addition we proposed a new approach to perform skull stripping by replacing the encoder part of UNet model by the pre trained deep network architectures. Pre trained deep models DenseNet121, MobileNetV2, InceptionResNetV2, VGG16 and VGG19 are used and their performance in the skull stripping on the same dataset are compared.