SPHARM-Mamba: Rotation-Invariant Multiscale Modeling for Brain Age Prediction

Official preview for the MICCAI 2026 paper SPHARM-Mamba: Rotation-Invariant Multiscale Modeling for Brain Age Prediction in the SPHARM-Mamba repository.

The proposed method performs cortical surface-based brain age prediction using spherical harmonic representations and Mamba-based sequence modeling. It captures multiscale spectral patterns on cortical surfaces while preserving rotation-invariant properties.

Description

Brain age prediction from cortical surfaces requires modeling both global cortical organization and local morphological variation. Existing surface-based deep learning methods often rely on local aggregation or patch-based tokenization. These designs can be sensitive to surface rotations and may require anatomical registration or extensive data augmentation.

SPHARM-Mamba addresses this limitation by representing cortical surface features in the spherical harmonic domain. Spherical harmonic convolution first extracts geometry-aware spectral features while preserving rotation-equivariant structure. The learned coefficients are then converted into rotation-invariant degree-wise descriptors through power aggregation within each harmonic degree.

The resulting descriptors form a natural coarse-to-fine sequence, where low harmonic degrees capture global shape patterns and high degrees represent increasingly localized cortical details. SPHARM-Mamba applies Mamba to this degree-ordered sequence to model cross-degree dependencies and integrate multiscale cortical information efficiently. This enables accurate and rotation-invariant brain age prediction without relying on patch partitioning or anatomical surface registration.

Getting Started

This section provides instructions for data preparation and training SPHARM-Mamba for cortical surface-based brain age prediction. Replace the example paths with your own directories.

Data Preparation

Convert FreeSurfer cortical surface features into icosphere-based representations (ico6).

python -m scripts.seg.prep \
    --data-dir "$SUBJECTS_DIR" \
    --out-dir "$DATA_DIR" \
    --hemi lh rh \
    --in-ch curv sulc \
    --ico 6

• $SUBJECTS_DIR: Directory containing FreeSurfer-processed subjects.
• $DATA_DIR: Directory for preprocessed surface data.

Chronological age labels should be prepared separately in $DATA_DIR/labels/ using the expected scalar label format (e.g., subject.lh.aug0.float.dat).

Training

Train SPHARM-Mamba using 5-fold cross-validation.

for fold in {1..5}; do
    python -m scripts.seg.train \
        --data-dir "$DATA_DIR" \
        --hemi lh rh \
        --in-ch curv sulc \
        --n-splits 5 \
        --fold "$fold" \
        --epochs 100 \
        --gpu 0 \
        --log-dir "./logs/fold${fold}" \
        --ckpt-dir "./checkpoints/fold${fold}"
done

For each fold, the model with the lowest validation MAE is saved as best_model_fold{fold}.pt. After training, the best checkpoint is automatically evaluated on the test set.

• test.log: Test MAE and evaluation metrics.
• test_age.log: Predicted and chronological ages for each test subject.

No separate inference script is required.