G-CNN: Adaptive Geometric Convolutional Neural Networks for MRI-Based Skull Stripping

Abstract:

Skull stripping in MRI-based brain imaging involves extraction of brain regions from raw images. While some Convolutional Neural Nets (CNNs) models have been successful in automating this process, the reliance on local textures can negatively impact model performance in the presence of pathological conditions such as brain tumors. This study presents a novel yet practical approach to offer supplementary texture-invariant spatial information of the brain as a geometric prior to enhance shape representations to further enable informed segmentation decisions. Our numerical results demonstrate that the new method outperforms competing algorithms in skull stripping tasks for both healthy and pathological inputs. These findings underscore the potential of incorporating geometric information in deep learning models to enhance the accuracy of brain segmentation.