Abstract: In recent years,the local-window based Self-Attention mechanism has gained prominence in vision tasks.However,due to the limited receptive field and weak modeling ability,it is not effective in dealing with complex data.The features in lung CT images are complex and diverse,including the shape,size and density of nodules,which bring challenges to mining the deep features in the data.To address these issues,this paper proposes a bi-directional multi-level interaction vision Transformer(Bi-MI ViT) backbone network that effectively integrates spatial and channel information through an innovative bi-directional multi-level interaction mechanism.This integration significantly improves the accuracy and comprehensiveness of feature extraction.Within the Transformer branch,we introduce an efficient cascaded group attention(CGA) strategy to enrich the diversity of attention head features and enhance the model’s ability to capture key information.Simultaneously,in the convolutional neural network(CNN) branch,we utilize a depth-wise and point-wise(DP) block structure along with point-wise convolution(PW) and depth-wise convolution(DW) to deeply mine local information and optimize model representation ability.Additionally,our establishment of a deep feature extraction(DFE) module enhances feature propagation and reuse while optimizing data utilization efficiency,leading to substantial performance improvement.Experimental results on both of the public COVID-CT dataset and private LUAD-CT dataset demonstrate that the proposed method outperforms the eight comparison methods in classification accuracy.

Key words: Lung CT images, Bi-directional multi-layer interaction, Convolutional neural network, Transformer, Classification