Abstract
Background: Hemorrhagic stroke is a life-threatening cerebrovascular disease, and early identification is crucial for timely clinical intervention. Microwave imaging is non-ionizing, portable, and low-cost, and thus has potential for pre-hospital and bedside screening; however, existing methods often suffer from limited reconstruction resolution, scarce data, and suboptimal information utilization when only a single modality is used. Methods: We propose a dual-channel, multi-input multimodal deep neural network for hemorrhagic stroke recognition, which jointly exploits complementary features from microwave images and time-domain waveforms and performs feature-level cross-modal fusion. A high-fidelity microwave brain simulation dataset is constructed for model training, and multiple temporal encoding strategies are systematically evaluated. Results: The proposed multimodal model achieves improved accuracy and stability compared with single-modality baselines and conventional approaches, demonstrating the benefit of cross-modal feature fusion for microwave-based hemorrhage recognition. Conclusions: Multimodal learning can enhance discrimination and robustness in microwave-based hemorrhage recognition, supporting its potential use for rapid, non-ionizing pre-hospital and bedside assessment.