Classification of Lung Adenocarcinoma Subtypes Based on ViT-DenseNet

Lung cancer is one of the most prevalent cancers with the highest mortality rate. And the classification of its adenocarcinoma subtypes is a challenge task for traditional CNN-based networks, due to such medical decision requires the comprehensively consideration of the complex spatial semantic information and the local texture information of adenocarcinoma tumors. Therefore, we propose an end-to-end dual stream network (ViT-DenseNet) to solve the problem of poor semantic representation of traditional CNNs with ViT’s strong global information modeling capabilities. Specifically, to cope with the negative fusion of the hybrid network, we propose a SRL method to improve the complementarity of the two feature extractors. Besides, the difference in inductive bias ability among sub-blocks is another critical factor incentive for negative fusion, we propose a MISM strategy to reduce the effect of such difference. Extensive experiments results demonstrate the effectiveness of SRL block and MISM strategy. And our approach also offers best average accuracy (85.88%) over other state-of-the-art models. In addition, by comparing visualization results, our proposed ViT-DenseNet dedicates more attention to the heterogeneous nuclei concerned by pathologists, thus offering stronger clinical interpretability.