A high-performance multimodal deep learning model for subtype classification of lung adenocarcinoma

Abstract This study focuses on the pathological subtype classification of lung adenocarcinoma (LUAD) in frozen sections and is the first to apply the state space model Vision Mamba to this task. We developed a fine-tuned multimodal MedMamba model that integrates histopathological images with non-image data. Through architectural optimization and multimodal fusion–including clinical variables and structured CT reports–the model achieved a well-balanced performance in classification accuracy, computational efficiency, and interpretability. In five-fold cross-validation, it achieved a recall of 92.0% and an F1-score of 91.6% for invasive adenocarcinoma (IAC) classification, as well as an accuracy of 89.7% and specificity of 92.0% for adenocarcinoma in situ (AIS) classification. On an external test set, the model maintained an accuracy of 84.2%, outperforming mainstream models such as InceptionV3, EfficientNetV2, and ResNet50. Further analysis revealed significant associations between tumor volume, radiologic appearance, and patient age with subtype classification, highlighting the potential of non-image modalities in diagnostic support. These findings demonstrate the practical value of the fine-tuned multimodal MedMamba model for intraoperative rapid pathology interpretation and clinical risk control.