Abstract: Due to the irregular shape and large difference of lung tumors in 3D medical images, the feature extraction of lesions is insufficient, and the recognition accuracy is not high, a 3D Cross-Modal ConvFormer is proposed. Firstly, three networks are utilized to learn the 3D PET, CT and PET/CT medical images. Secondly, a ConvFormer model is designed to fuse global and shallow local features, while self-correcting convolution expands the receptive field for better lesion extraction. Finally, a dual-branch cross-modal feature interaction block is designed to enhance cross-modal features and capture 3D multimodal details. This module uses two global attention mechanisms to improve the extraction of cross-modal and global-local information. The experiments use a 3D multimodal lung tumor dataset with 3 173 patients. With optimized parameters and computation time, the 3D-CConvFormer achieves an accuracy of 89.25% and an AUC of 88.74%, providing reliable computer-aided diagnosis for 3D multimodal lung tumor.