The application belongs to the field of diagnosis and treatment, and particularly relates to a
lung cancer EGFR mutation non-invasive prediction model based on unsupervised decoupling
and gate fusion. A multi-
modal fusion model is constructed based on
enhanced CT images and clinical data of non-
small cell lung cancer patients. The application adopts a
mask image modeling task, and autonomously learns morphological and texture primitive features of tumors from three-dimensional
enhanced CT images without
EGFR mutation labels. The application overcomes the dependence on a large amount of
labeled data, enables the model to learn visual representations with high discriminability and generalization ability from unlabeled data, and significantly improves feature expression capability. A multi-
modal fusion architecture based on a gate mechanism is designed, which can dynamically adjust the contribution weights of deep features, imaging features and
clinical information in different samples. Personalized
feature fusion for different cases is realized, the limitations of traditional static fusion methods are overcome, and complementary information between multi-source heterogeneous data is fully mined.