Fig. 1: Overall architecture of kV2CTConverter.
a Workflow of the proposed method. The raw kV images were augmented by GRSS to get adequate samples for model training. Then the processed images simultaneously went through dual models (i.e., primary model and secondary model) to generate the whole CT and the fractional CT that covered only the head region, respectively. Lastly, the full-size synthesized CT was achieved by overlaying and concatenating the outputs from two models according to their spatial relationship. b The model structure of both primary and secondary model. c. The details of the hierarchical ViT blocks in the encoder Ek. d The details of the hierarchical ViT blocks in the decoder Dr. e The detailed illustration of the window-based Multi-Head Attention (W-MHA), the tokenized patches were first spat to nW non-overlapped windows of a size of w × w and the attention was only calculated on the windows instead of the whole inputs.
