EM-AWFNet: dual-branch adaptive weight fusion for lung cancer PET/CT classification

Lung cancer (LC) is a malignant tumor that poses a serious threat to human health. Positron emission tomography/computed tomography (PET/CT) imaging, which integrates both anatomical structure and metabolic activity, plays a crucial role in the auxiliary diagnosis of LC. Despite its advantages, the automated classification of LC subtypes is challenged by the overlapping imaging features among different subtypes and the complex morphological characteristics of the lesions. To address the aforementioned challenges, this paper proposes EfficientNet-MobileViT adaptive weighted fusion network, a dual-branch adaptive weight fusion network based on deep learning for LC classification using fused PET/CT images. The method employs a pretrained efficientnet-B1 as the local path to extract detailed structural features, while MobileViT serves as the global path to capture long-range contextual semantic information. A Transformer module is simultaneously introduced to facilitate cross-branch feature interaction. Furthermore, an adaptive weight fusion mechanism is designed to dynamically adjust the contribution of each feature branch during the classification process. Experimental results demonstrate that the proposed method achieves a classification accuracy of 0.9389 on the Lung-PET-CT-DX dataset for the three-class classification task of lung adenocarcinoma, small cell carcinoma, and squamous cell carcinoma. It provides new insights for research on intelligent analysis and computer-aided diagnosis of LC using fused PET/CT images.