An efficient R-Transformer network with dual encoders for brain glioma segmentation in MR images

Abstract

Gliomas are the most prevalent and destructive forms of primary brain tumors with a high mortality rate. Magnetic resonance imaging is extensively employed in the examination of gliomas. Segmenting brain glioma on magnetic resonance images has significant clinical value. However, the blur boundary of gliomas, variability in the shape, location, and size make segmentation extremely challenging. In this paper, we propose a new method to segment gliomas from three-dimensional brain magnetic resonance images accurately. We propose an efficient R-Transformer network with dual encoders (ERTN) to achieve precise segmentation by innovatively combining R-Transformer with U-Net. Specifically, ERTN constructs a feature branch and a patch branch, capturing complex semantic features and global context information. Moreover, features generated from the feature branch and patch branch are up-sampled and combined with low- and high-resolution CNN features in the decoder to enable precise localization. At last, ERTN employs the ranking attention mechanism in Transformer (R-Transformer), which helps the model focus on helpful information to improve training efficiency and reduce computation cost. Experiments on the 2017 BRATS dataset prove that ERTN achieves satisfactory performance, with a Dice similarity coefficient of 83.20%, 77.93%, and 72.59% on the whole tumor, tumor core, and enhanced tumor segmentation. For the Hausdorff distance index, we obtained the scores of 5.30, 4.60, and 5.50 for the whole tumor, tumor core, and enhanced tumor, respectively. Our results suggest that ERTN improves the segmentation accuracy and reduces computation cost, which performs better than the existing convolution- and transformer-based methods.