Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
BrainODE: Neural Shape Dynamics for Age- and Disease-aware Brain Trajectories
Authors: Wonjung Park, Suhyun Ahn, Maria Hernandez, Susana Maniega, Jinah Park
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that Brain ODE outperforms time-aware baselines in predicting future brain shapes, demonstrating strong generalization across longitudinal datasets with both regular and irregular time intervals. ... In this section, we demonstrate the comparison of Brain ODE with baseline methods for time series and then discuss the effect of our proposed components. We conduct experiments using four different longitudinal brain MRI datasets of normal cognition and Alzheimer s disease (AD)... |
| Researcher Affiliation | Academia | Wonjung Park KAIST EMAIL Suhyun Ahn KAIST EMAIL Maria C. Valdes Hernandez The University of Edinburgh EMAIL Susana Muñoz Maniega The University of Edinburgh EMAIL Jinah Park KAIST EMAIL |
| Pseudocode | Yes | Algorithm 1 Brain ODE training process ... Algorithm 2 The pseudo-cognitive status embedding |
| Open Source Code | Yes | Our source code, including data preprocessing scripts and model configurations, is publicly available at https://github.com/PWonjung/Brain ODE |
| Open Datasets | Yes | We conduct experiments using four different longitudinal brain MRI datasets of normal cognition and Alzheimer s disease (AD), comprising both regular and irregular time intervals: for regular intervals, the Lothian Birth Cohorts 1936 (LBC1936) [9] and the Australian Imaging, Biomarker and Lifestyle (AIBL) datasets [10], each collected at fixed 3 and 1.5 year intervals, respectively; for irregular intervals, the Alzheimer s Disease Neuroimaging Initiative (ADNI) [12] and the Open Access Series of Imaging Studies (OASIS) datasets [20], with varying acquisition intervals reflecting real-world clinical settings. |
| Dataset Splits | Yes | Given the varying number of longitudinal observations N ranging from 2 to 5 across the datasets, we formulate the brain disease progression modeling into two tasks: (1) predicting the latest shapes for each subject using four observed time points (4-shot prediction), and (2) predicting the latest shapes from a single observation point (1-shot prediction). ... The validation dataset composition is Table 6, used for the quantitative results in Table 1 and Table 2. Specifically, we selected subjects whose longitudinal observations are more than four times for the 4-shot prediction. |
| Hardware Specification | No | In contrast, Brain ODE employs mesh representations and PCA-based 3D shape parameterization, which significantly reduces computational requirements (~375 MB for training) while achieving high-precision shape modeling. |
| Software Dependencies | No | Brain ODE is trained for 100 epochs using the Adam W optimizer with a learning rate of 0.0005. For solving the neural ordinary differential equations, we employ the fourth-order Runge-Kutta (RK4) integration method. ... As part of this pipeline, we use Synth Seg [1] to segment brain regions from MRI scans. |
| Experiment Setup | Yes | Brain ODE is trained for 100 epochs using the Adam W optimizer with a learning rate of 0.0005. For solving the neural ordinary differential equations, we employ the fourth-order Runge-Kutta (RK4) integration method. ... We empirically set hyperparameters for each loss as {λcf, λpm, λvert, λnorm, λedge, λcons, λlap} = {0.5, 3, 1, 1, 1500, 1, 5}. ... The temporal input t is derived by normalizing the age of subjects. Specifically, we select subjects aged between 65 and 95 years and linearly normalize their age (i.e., t = (age 65)/30 such that t [0, 1]). |