Predicting Brain Morphometry with MT-GNN: Continuous-Time Mesh Evolution via Graph-Based Metric Tensor Embeddings
Original title:Predicting Brain Morphometry with MT-GNN: Mesh Evolution in Continuous Time with Graph-Based Metric Tensor Embeddings
The paper introduces MT-GNN, a continuous-time longitudinal mesh predictor that forecasts each vertex’s first fundamental form rather than directly predicting vertex displacements. A Fourier encoding represents the lead time, while a differentiable As-Rigid-As-Possible solver decodes the predicted metric tensor into a surface. End-to-end training uses rigid-aligned vertex error. On 14 subcortical structures from ADNI, MT-GNN reportedly outperforms the temporal-mean baseline, DCM geodesic shape regression, and TransforMesh at every evaluated horizon, with an average vertex-error improvement of 2.29% over the temporal mean.
Why it's worth reading
The reported ADNI results show consistent gains across all 14 structures and prediction horizons, making the paper timely for assessing whether intrinsic-geometry modeling can improve disease-progression forecasting and clinical-trial enrichment.