Enhancing brain boundary segmentation with CM-BET: A cascaded model for high precision in brain extraction

Gürkahraman K., Karakış R., Yelkuvan A. F.

JOURNAL OF THE FACULTY OF ENGINEERING AND ARCHITECTURE OF GAZI UNIVERSITY, cilt.41, sa.2, ss.995-1009, 2026 (SCI-Expanded, Scopus, TRDizin)

Özet

Brain extraction is a critical task in neuroimaging that enables the separation of brain from surrounding tissues and directly affects the performance of subsequent analyses. Deep learning (DL)-based approaches have shown promising results in brain extraction, where accurate segmentation of brain boundaries plays a crucial role in model performance. In this study, a two-stage 3D DL model named CM-BET is proposed. In the first stage, the MRS-UNet generates a grayscale brain output by emphasizing high-frequency edge information through sharpening layers integrated into the encoder blocks of UNet. In the second stage, the MR-UNet takes this sharpened image as input and performs more precise brain segmentation to generate a binary brain mask. Residual connections adopted from ResNet50 were incorporated into the latent layer to preserve low-gradient boundary information of the brain. In addition, a hybrid loss function containing focal loss was developed to address the class imbalance problem commonly encountered in MRI slices. Evaluations of seven public datasets demonstrate that CM-BET outperforms DL-based methods in literature and widely used segmentation models such as UNet, UNet++, and DeepLabv3+. The proposed model achieved minimum-maximum dice coefficient values of 0.9809-0.9984, sensitivity values of 0.9831-0.9991, and specificity values of 0.9958-0.9992 across all datasets, outperforms DL-based studies and UNet variants.