Alzheimer's diagnosis moves to two brain graphs: AD versus CN accuracy rises from 82.06% to 89.90%
Synopsis
Structural MRI brain graphs no longer rely on a single predefined connection rule. MVMGNN treats brain regions as nodes carrying radiomics features, builds a spatial-neighbor graph and a feature-similarity graph, filters redundant features and weak connections inside each view with a joint node-edge mask, and fuses the two views through patient-level gated fusion. On the ADNI cohort, AD versus CN accuracy rises from 82.06% for the best comparison method, RSEA-MVGNN, to 89.90%, and F1 from 75.02% to 86.24%; gains on the harder AD-MCI and CN-MCI tasks are smaller.
Interpretation
Staging Alzheimer's disease from structural MRI now learns on two complementary brain-region graphs at once. Earlier brain-graph methods mostly built edges from a single predefined relationship, such as spatial location or morphological distance, so they could not use inter-regional relationships from different perspectives together. On the ADNI AD versus CN task, MVMGNN reaches 89.90% accuracy and 86.24% F1, against 82.06% and 75.02% for RSEA-MVGNN, the best-performing comparison method on that task.
Each view is filtered before graph convolution propagates information. Concatenating multi-view brain networks directly brings in redundant radiomics features and weakly relevant connections that interfere with message passing. On AD versus CN, the basic dual-view model reaches 82.78% accuracy and 76.57% F1; adding the node-edge mask raises these to 88.92% and 85.26%.
The two view representations are combined with per-subject weighting rather than fixed weights. Fixed or simple linear weighting cannot adjust each view's contribution to an individual sample. On AD versus CN, adding cross-view gated fusion to the masked model lifts accuracy and F1 further to 89.90% and 86.24%, and specificity from 89.42% to 91.69%.
The same modeling still yields usable classification on two public cohorts beyond ADNI. ADNI training and test splits are made at the subject level, so all scans from one subject enter either the training set or the test set. On AIBL and OASIS, which were not used in training or testing, AD versus CN accuracy is 87.46% and 80.38% respectively.
Perspective
The demonstrated scope is structural MRI alone: 908 subjects from the ADNI cohort split at the subject level into 80% training and 20% testing, plus external validation on the public AIBL and OASIS cohorts. The intended readers are researchers and engineers working on neuroimaging graph learning and computer-aided diagnosis.
Mask weights reflect the model's reliance on brain-region features and connections during classification and cannot be read directly as causal biomarkers of disease mechanism. Multi-view construction relies on predefined spatial-proximity and feature-similarity relationships, whose optimality still requires validation. External validation is limited by the availability of independent cohorts and by differences in sample composition and imaging characteristics.