What problem does it solve? Applying deep learning to whole-brain MRI volumes requires a well-defined model, a strict input contract, and reproducible cross-validated training. This Skill provides a canonical compact residual 3D CNN (VoxelCNN3D) with a fixed NPZ input format and checkpointing so voxel-level classification and regression experiments run consistently. ## Core Features & Use Cases - Voxel-level classification and regression: Train a residual 3D CNN on structural MRI, fMRI summaries, statistical maps, or other aligned volumetric data. - Cross-validated training with provenance: Produces predictions.csv, fold_assignments.csv, metrics.json, per-fold checkpoints, and run manifests for reproducibility. - Strict input contract: Accepts a single .npz file with X [subjects, channels, depth, height, width], y, and optional subject_id, enforcing consistent orientation, voxel size, and normalization. - Use Case: A researcher with skull-stripped, normalized sMRI volumes for 200 subjects wants to classify diagnosis groups; they package the volumes into an NPZ and run the reference training script with 5-fold cross-validation on a CUDA device. ## Quick Start Ask the AI to train the cnn3d model on your prepared volumes.npz file for classification with 5-fold cross-validation and save outputs to a run directory.
