Automated and Robust Brain Skull Stripping using Optimized Pre-processing and a Refined Residual U-Net Framework

Abstract

 In this study, a method for skull stripping referred to as the 3D Enhanced Residual U-Net is introduced. The approach combines the traditional U-Net with an enhancement mechanism designed to improve both the effectiveness and processing speed of the U-Net. An anisotropic diffusion filter (ADF) reduces noise in MRI images while maintaining the edges of present objects. This is accompanied by skull stripping to eliminate non-brain matter and contrast enhancement to elevate the visual quality. The architectural adaptations allow for rapid and stable training. As brain images vary significantly from subject to subject, a deep learning approach will account for these differences leading to consistent skull stripping results. Neurofeedback Skull Stripping (NFBS) dataset was used for the proposed model formulation. The results from the experiments show that the proposed approach is effective and practical compared to previous methods. This method obtained an impressive sensitivity rate of 0.9974, DSC of 1.0000, a specificity of 0.9983, an IOU of 0.9831, an accuracy percentage of 0.9969, and a precision of 0.9961, showing an actual ability to differentiate the distinct parts of the brain and the skull.