부산대학교 도서관

Image segmentation is an essential aspect of image processing, where regions of interest (ROI) are identified by employing feature descriptors like edges, color, and texture. Medical images often contain unwanted segments, necessitating preprocessing steps to extract valuable information. In the realm of medical image segmentation, skull-stripping in 3D volumetric Brain MRI images holds significant importance. This study delves into supervised learning techniques to extract crucial brain features from non-brain tissues. The approach involves a modified U-net that operates on 2D slices of the brain MRI, yielding a skull-stripped image with superior accuracy compared to existing methods. The architecture incorporates Adaptive Polyphase Pooling layers, enhancing the architecture’s robustness and the model’s generalization. These Adaptive Polyphase Pooling layers ensure shift-invariance in the predicted output, maintaining consistency even if the single view of the Brain MRI shifts along the two axes in the plane.