Abstract

Brain extraction refers to the process of removing non-brain tissues in brain scans and is one of the initial pre-processing procedures in neuroimage analysis. Since errors produced during this process can be challenging to amend in subsequent analyses, accurate brain extraction is crucial. Most deep learning-based brain extraction models are optimised on performance,leading to computationally expensive models. Such models may be ideal for research; however, they are not ideal in a clinical setting. In this work, we propose a new computationally efficient 2D brain extraction model, named RGU-Net. RGU-Net incorporates Ghost modules and residual paths to accurately extract features and reduce computational cost. Our results show that RGU-Net has 98.26% fewer parameters compared to the original U-Net model, whilst yielding state-of-the-art performance of 97.97 ± 0.84% Dice similarity coefficient. Faster run time was also observed in CPUs which illustrates the model’s practicality in real-world applications.