Can ML models trained on raw neonatal MRI acquired at term-equivalent age reliably learn high-dimensional imaging features that predict later cognitive outcomes in extremely preterm children, and do these models outperform traditional regression approaches based on predefined imaging metrics? While the framework tested so far successfully addresses individual spatial characterization, several technical extensions are planned to enhance its clinical utility: • Hyperparameter Tuning of the SSL Encoder via NAISS: The immediate next computational phase will involve deep hyperparameter tuning of the 3D Swin-UNETR backbone (exploring alternative masking ratios, patch dimensions, and learning rate sched- ules). • Long-Term Cognitive Outcome Linkage: A prospective extension will link the individual parcellation-anchored deviation statistics directly to long-term cognitive and functional performance scores (e.g., Bayley-III scores) collected at school-age follow-ups. • Leave-One-Out Cross-Validation (LOOCV) Null: To address the current in-sample bias (where the same 21 controls define the null and train the autoencoder), a LOOCV loop can be implemented. • Multi-Channel Reconstruction Autoencoders: To separate the confounding factors of cortical thickness, surface folding, and tissue signal intensity within the reconstruction residuals, future iterations will expand the autoencoder from a single-channel T1-weighted input to a multi-channel framework incorporating T2-weighted and diffusion-weighted imaging metrics.