Transformer-based multimodal fusion model predicts lymph node metastasis in hepatic alveolar echinococcosis patients: A multicenter study.

OBJECTIVES

To develop a CT-based multimodal transformer model to precisely predict lymph node (LN) metastasis in hepatic alveolar echinococcosis (HAE) patients.

METHODS

A total of 318 HAE patients from three centers were allocated to a training set, an internal validation set, and two external validation sets. Radiomics, 3D deep learning (3DDL), and 2D deep learning (2DDL) features were retrieved from contrast-enhanced CT images of the hepatic hilar LN. Random forest models were constructed utilizing various features. Ultimately, we developed and assessed a transformer-based multimodal fusion model. Model performance was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve and decision curve analysis (DCA).

RESULTS

Compared with both the radiomics and the 2DDL models, the 3DDL model exhibited enhanced discriminative ability for LN status. The transformer model achieved the highest AUC (95%CI) of 0.951 (0.898-1.000), 0.927 (0.850-1.000), and 0.933 (0.847-1.000) for the three validation sets. DCA revealed that the transformer model produced the greatest net clinical advantage.

CONCLUSION

This study innovatively constructed a transformer-based multimodal fusion model, providing a practical and reliable tool for predicting LN metastasis in HAE patients. More importantly, this model provides a foundation for guiding LN dissection in HAE patients and is readily applicable in clinical settings.