Segmentation Method of Magnetoelectric Brain Image Based on the Transformer and the CNN

To address the problem of a low accuracy and blurred boundaries in segmenting multimodal brain tumor images using TransBTS network, 3D BCS_T model incorporating channel space attention mechanism is proposed. Firstly, hierarchy increased to obtain more local feature information, residual basis blocks are added reduce loss. Secondly, downsampling incorporated into hybrid enhance critical region information extraction. Finally, weighted cross-entropy loss generalized dice employed solve inequality sample categories. The experimental results show that whole WT, core TC, enhanced ET improved by an average 2.53% evaluation metric Dice similarity coefficient, compared with network shortened 3.14 Hausdorff distance 95. Therefore, can effectively improve segmentation boundary clarity both categories small areas.