Bidirectional generative diffusion model with cascaded symmetric attention for brain connectivity-to-connectivity translation

Understanding the complex relationship between structural and functional connectivity is fundamental to discovering the pathogenesis of brain diseases. However, it is challenging to establish nonlinear connective relationships between structural and functional modalities. In this study, a novel bidirectional generative diffusion model (BGDM) is proposed to construct bidirectional mapping for connectivity-to-connectivity translation. The proposed BGDM is an AIGC framework to learn bidirectional translation between structural and functional connectivity. By designing the cascaded symmetric attention module, the BGDM can learn multi-channel and cascaded edge features to estimate diffusion noise, which improves computational efficiency and enhances the model’s ability to predict brain connectivity patterns. Furthermore, the connectivity noise-balanced loss is devised to adaptively adjust the importance of individual noisy connections and ensure more accurate and reliable translations between diverse brain connectivity representations. We demonstrate the effectiveness of BGDM through comprehensive experiments on the HCP datasets. Our model achieves the best translation performance with the mean MAE/SSIM of 0.065/0.835 for SC-to-FC translation, and the mean MAE/SSIM of 0.050/0.932 for FC-to-SC translation. Compared with the state-of-the-art method, the results of our model show improvements of 0.014 (MAE) and 0.06 (SSIM) for the SC-to-FC task, 0.014 (MAE) and 0.04 (SSIM) for FC-to-SC task. Our model can generate incomplete multimodal brain networks, which can be used to improve the accuracy of brain disease diagnosis and provide clinicians with disease-related biomarkers. Moreover, the proposed model has the potential to bridge the gap between structural connectivity and functional connectivity, offering new opportunities for understanding the brain’s working mechanisms and revealing brain disease’s pathogenesis.