Objective Stroke poses a heavy burden due to its high mortality and morbidity rates. Accurate and real-time detection of lesions is pivotal for prompt clinical intervention and favorable prognosis. Electrical impedance tomography (EIT) and microwave tomography (MWT) have emerged as compelling alternatives for stroke screening, owing to their non-ionizing, non-invasive and portable nature. EIT provides information on tissue conductivity, and MWT offers high sensitivity to changes in dielectric properties. However, single-modality imaging is inherently limited, EIT suffers from low sensitivity to deep-seated tissues and severe ill-posedness of inverse problems, whereas MWT is challenged by strong nonlinearity in inverse scattering and susceptibility to modeling errors. Consequently, the clinical utility of standalone EIT or MWT for stroke diagnosis remains constrained by poor spatial resolution and imaging artifacts. Methods To improve the accuracy and robustness of stroke imaging, a dual-modality fusion conditional denoising diffusion probabilistic model (DM-DDPM) was proposed for high-precision dual-modality image reconstruction. A dual-encoder network with a symmetric architecture and independently trained parameters was constructed to extract heterogeneous features separately from EIT boundary voltage measurements and MWT scattered field signals. Attentional feature fusion (AFF) is employed to integrate complementary information from the two modalities adaptively, generating robust fused priors that suppress redundant noise while preserving key physical characteristics. Subsequently, the fused priors are embedded into a Transformer-based diffusion model via a cross attention mechanism to guide the reverse denoising process. This approach effectively reduces artifacts and enhances the stability of conductivity distribution reconstruction. Time step embedding is introduced to enable the network to perceive the diffusion stage and further improve the accuracy of noise prediction. Results Simulated experiments demonstrated that DM-DDPM significantly outperforms single-modality and multi-modality networks under various noise levels. A head model simulation dataset was constructed based on COMSOL Multiphysics, and tests were carried out under 50 dB, 40 dB and 30 dB signal-to-noise ratio levels. At 30 dB, the average relative error (RE) was below 0.20, while the structural similarity index (SSIM) and correlation coefficient (CC) remained above 0.90 and 0.89, respectively. Compared with single-modality and multi-modality networks, artifacts were significantly reduced, lesion edges were clearer, and localization was more accurate. The model maintains high reconstruction quality and strong robustness for single, double, and triple lesions simultaneously. Furthermore, physical experiments were conducted using a 16-electrode EIT system and a 16-antenna MWT system with asynchronous data acquisition. These experiments confirmed the feasibility of the method in real-world scenarios and demonstrated that it can robustly reconstruct simulated lesions despite environmental interference and measurement noise, validating its reliability for practical clinical applications. Conclusion The proposed method effectively combines complementary dual-modality information with a conditional diffusion model. Low accuracy and poor noise resistance in single-modality imaging were effectively addressed, while the noise amplification issue caused by direct multimodal data fusion was avoided. The proposed algorithm exhibits strong anti-noise interference ability and high imaging stability in both simulation and physical experiments. Precise localization of stroke lesions with different quantities was achieved, providing a high-precision, and practical technical support for clinical stroke detection.