Abstract
Accurate and early diagnosis of esophageal cancer (EC) remains challenging due to heterogeneous imaging characteristics and the complexity of integrating multimodal data. This study proposes a multimodal-multiscale hybrid fusion network (MHF-Net), which combines domain-driven handcrafted features (HFs) and deep multiscale features (DFs) from multimodal data for the automated diagnosis of EC. MHF-Net uses a dilated-inception block to extract multiscale representations at varied dilation rates and concatenates these DF with domain-specific features via dense connections. A convolutional block attention module further refined spatial and channel-wise features, leveraging global average pooling. The grasshopper optimization algorithm optimized fusion weights and hyperparameters, enhancing hybrid feature integration and overall model robustness. Evaluated across five diverse datasets, MHF-Net achieves state-of-the-art performance (accuracy = 0.970 ± 0.014; F1 score = 0.960 ± 0.017). This study demonstrates clinical applicability with strong potential for future enhancement through the integration of multimodal biomarkers for enhanced diagnostic accuracy.