Introduction: Moyamoya disease is a progressive cerebrovascular disorder. Revascularization is an effective procedures to reduce the risk for stroke. While the goal is to restore normal cerebral blood flow, predicting the outcome of this procedure is challenging. AI has shown great promise in predicting and synthesizing imaging results. Here, we a present a novel strategy using AI to predict post-operative CBF from pre-operative MRI scans in moyamoya. Methods: A conditional latent diffusion model based on the MONAI framework was used to train the prediction model. The model included variational autoencoders with 2D spatial dimensions and a diffusion process operating in the latent space. A total of 124 retrospective paired pre- and post-operative CBF images derived from ASL MRI were included for training the diffusion model. Preoperative CBF images were used as conditional inputs to predict 6-month postoperative outcomes as model outputs. A total of 25 cases were reserved for validation. Training utilized a multi-component loss function: perceptual loss to preserve anatomical features, adversarial loss to ensure realistic image quality, and mean squared error to ensure pixel-wise accuracy. We assessed prediction accuracy using structural similarity index (SSIM), mean squared error (MSE), and voxel-wise correlation coefficients. Results: Figure 1 shows the prediction of an example patient. Overall, the model successfully predicted the improved CBF. The model generated spatially coherent predictions that preserved anatomical structure and maintained reasonable intensity distributions. In terms of quantitative results, validation on the independent test set demonstrated mean SSIM scores of 0.77 and correlation coefficients of 0.63 between predicted and actual postoperative images. MSE values were low and averaged 0.14. Discussion: This novel application of conditional latent diffusion models demonstrates clinically relevant accuracy in predicting postoperative CBF changes in Moyamoya disease. The model offers potentially patient-specific predictions that could enhance patient selection and procedural planning. The approach advances precision medicine in cerebrovascular surgery by leveraging generative AI for medical imaging applications. While the approach demonstrates the feasibility of applying generative AI to CBF prediction, further optimization is needed to improve voxel-wise correspondence with postoperative changes, which is an area of our ongoing investigation.
Zhao et al. (Thu,) studied this question.