Fig. 9: The network structure of MultiChannel 2.5D CrossFormer model.

The red rectangle and the blue rectangle represent the slices obtained from the T1CE and T2FLAIR sequences, respectively, while the purple rectangle indicates the smallest rectangle that covers the tumor area in all six slices (i.e., the defined ROI as a 2.5D input). The input size is H0 × W0. The size of the feature maps at each stage is indicated at the top, and Stage-i specifically comprises a CEL (convolutional embedding layer) and ni CrossFormer blocks. The number within the CEL denotes the kernel size utilized for patch sampling.