Predicting the progression from Mild Cognitive Impairment(MCI)to Alzheimer's Disease(AD)is a critical challenge for enabling early intervention and improving patient outcomes.While longitudinal multi-modal neuroim...Predicting the progression from Mild Cognitive Impairment(MCI)to Alzheimer's Disease(AD)is a critical challenge for enabling early intervention and improving patient outcomes.While longitudinal multi-modal neuroimaging data holds immense potential for capturing the spatio-temporal dynamics of disease progression,its effective analysis is hampered by significant challenges:temporal heterogeneity(irregularly sampled scans),multi-modal misalignment,and the propensity of deep learning models to learn spurious,noncausal correlations.We propose CASCADE-Net,a novel end-to-end pipeline for robust and interpretable MCI-to-AD progression prediction.Our architecture introduces a Dynamic Temporal Alignment Module that employs a Neural Ordinary Differential Equation(Neural ODE)to model the continuous,underlying progression of pathology from irregularly sampled scans,effectively mapping heterogeneous patient data to a unified latent timeline.This aligned,noise-reduced spatio-temporal data is then processed by a predictive model featuring a novel Causal Spatial Attention mechanism.This mechanism not only identifies the critical brain regions and their evolution predictive of conversion but also incorporates a counterfactual constraint during training.This constraint ensures the learned features are causally linked to AD pathology by encouraging invariance to non-causal,confounder-based changes.Extensive experiments on the Alzheimer’s Disease Neuroimaging Initiative(ADNI)dataset demonstrate that CASCADE-Net significantly outperforms state-of-the-art sequential models in prognostic accuracy.Furthermore,our model provides highly interpretable,causally-grounded attention maps,offering valuable insights into the disease progression process and fostering greater clinical trust.展开更多
文摘Predicting the progression from Mild Cognitive Impairment(MCI)to Alzheimer's Disease(AD)is a critical challenge for enabling early intervention and improving patient outcomes.While longitudinal multi-modal neuroimaging data holds immense potential for capturing the spatio-temporal dynamics of disease progression,its effective analysis is hampered by significant challenges:temporal heterogeneity(irregularly sampled scans),multi-modal misalignment,and the propensity of deep learning models to learn spurious,noncausal correlations.We propose CASCADE-Net,a novel end-to-end pipeline for robust and interpretable MCI-to-AD progression prediction.Our architecture introduces a Dynamic Temporal Alignment Module that employs a Neural Ordinary Differential Equation(Neural ODE)to model the continuous,underlying progression of pathology from irregularly sampled scans,effectively mapping heterogeneous patient data to a unified latent timeline.This aligned,noise-reduced spatio-temporal data is then processed by a predictive model featuring a novel Causal Spatial Attention mechanism.This mechanism not only identifies the critical brain regions and their evolution predictive of conversion but also incorporates a counterfactual constraint during training.This constraint ensures the learned features are causally linked to AD pathology by encouraging invariance to non-causal,confounder-based changes.Extensive experiments on the Alzheimer’s Disease Neuroimaging Initiative(ADNI)dataset demonstrate that CASCADE-Net significantly outperforms state-of-the-art sequential models in prognostic accuracy.Furthermore,our model provides highly interpretable,causally-grounded attention maps,offering valuable insights into the disease progression process and fostering greater clinical trust.
