文献类型：期刊文献

英文题名：RAF-MNet: A Role-Aware Frequency-Enhanced Multi-modal Network for Stroke Prognosis Prediction

作者：Wang, Ziying[1]; Zhu, Hongqing[1]

机构：[1] East China University of Science and Technology, School of Information Science and Engineering, Shanghai, China

年份：2026

外文期刊名：DLCV 2026 - 2026 IEEE 3rd International Conference on Deep Learning and Computer Vision

收录：EI(收录号：20263421337843)

语种：英文

外文关键词：Activation analysis - Computation theory - Decision making - Diagnosis - Forecasting - Frequency modulation - Medical imaging - Modal analysis - Semantics

摘要：Early prognosis prediction is important for individualized stroke management and treatment decision-making. However, stroke outcomes are affected by multiple factors, including lesion-related imaging patterns, neurological status, laboratory indicators, and clinical risk factors, making single-modality prediction insufficient. Existing multi-modal methods often adopt symmetric fusion strategies and rarely consider the different roles and potential optimization conflicts among modalities. This paper proposes RAF-MNet, a role-aware frequency-enhanced multi-modal network for stroke prognosis prediction. The proposed model jointly exploits NCCT imaging data, automatically derived textual descriptions, and structured clinical variables for outcome prediction. A frequency-aware imaging branch based on 3D discrete wavelet transform is designed to enhance lesion-related structural and high-frequency representations while suppressing noise-sensitive responses. A modality interaction module further fuses imaging evidence with textual semantics. To regulate modality-specific optimization, a role-aware asymmetric gradient optimization (RAGO) strategy is introduced. Experiments on a public dataset show that RAF-MNet outperforms representative single-modal and multi-modal baselines, achieving an AUC of 0.898, accuracy of 0.861, F1-score of 0.889, and recall of 0.898. Interpretability analysis using class activation maps and SHAP further demonstrates the model's ability to provide meaningful evidence from both imaging and tabular perspectives. ? 2026 IEEE.