Image-Text Fusion Network for Incomplete Multi-Label Brain Cognitive Risk Prediction

White matter hyperintensities (WMH) play a significant role in predicting cognitive risk in the human brain. Both their location and size can affect normal cognitive functions. Convolutional Neural Networks (CNNs) are widely used in risk prediction, but they are limited by their inherent mechanisms. For example, pooling layers can lead to a lack of local and global semantic correlation in images, which constrains prediction and classification performance. In this paper, we propose an Image-Text Fusion Network architecture that combines WMH images, patient physiological indicators, and clinical symptoms. This architecture extracts and semantically matches features from both lesion and clinical symptom modalities to generate a supervisory vector that can enhance the semantic expression of the classification network. Furthermore, we introduce a Multi-Branch Transformer module to improve the fusion of feature maps from different branches. This module emphasizes long-range dependencies between features from different channels and supervisory vectors at different scales. The final experimental results achieve an accuracy of 0.956 and an F1-Score of 0.836, outperforming existing classification networks. These results demonstrate the algorithm's potential to aid clinicians in diagnosis.