This paper proposes a method for the subtyping and diagnosis of lung cancer using a deep learning approach, with the aim of improving the performance of deep modelling of lung cancer subtypes using PET/CT images. METHODS: A novel approach to learning representations for lung cancer subtyping is proposed, whereby convolutional, state space, and attention layers are fused to model semantic features across different lung cancer subtypes in PET/CT images. The resulting representations are subsequently employed in the downstream task of lung cancer subtyping. RESULTS: the proposed method outperforms state-of-the-art models in the field of computer vision, including VGG19, ResNet34, ViT-b16, Swin-T, MAE, and others, on the publicly available Lung-PET/CT-Dx dataset. Our method achieved an average AUC score improvement of 6.9%, 5.4%, 4.1%, 2.7%, and 1.6% and an average F1 score improvement of 28.4%, 25.4%, 18.2%, 9.3%, and 3.3%, respectively. CONCLUSION: The method proposed in this paper, markedly enhances the classification efficacy of lung cancer subtypes utilizing PET/CT images. This illustrates the potential of this method to serve as an efficacious approach for the accurate diagnosis of lung cancer subtypes using PET/CT images in clinical practice.