Pipeline: From MRI to Latent Signatures

Multi-Dataset Robustness

The NeuroScope pipeline converts structural T1-weighted MRI and clinical variables into subject-level representations of neurodegenerative change. Our methodology is validated across OASIS-1 and ADNI-1 using standardized feature extraction to ensure cross-dataset compatibility.

1. Harmonization

Preprocessing & site-alignment

• Bias field correction & intensity normalization

• Skull stripping & template-space (MNI) alignment

• Resolution matching across multi-site datasets

• Handling label definition shifts (CDR vs. DX_bl)

2. Feature Extraction

ResNet18 + Anatomical

• 512D MRI embeddings via pre-trained ResNet18

• Regional brain volumes (Hippocampus, Ventricles)

• Global atrophy & Intracranial Volume (eTIV)

• Clinical covariates (Age, Sex, Education)

3. Multimodal Fusion

Late & Attention Fusion

• Modality-specific MLP encoders

• Late fusion (logit averaging/concatenation)

• Cross-modal attention mechanisms

• Explicit age modeling as a confounder

Experimental Consistency

Ensuring publication-grade results

To ensure results are truly robust, we utilize the exact same architectures and hyperparameters for both in-dataset training and cross-dataset evaluation.

MRI-Only models serve as the baseline, with Multimodal (Late/Attention) fusion architectures evaluating the additive value of clinical signal in predicting early cognitive decline.

Robustness philosophy

Beyond internal validation

Deep learning models often overfit to center-specific acquisition protocols. By testing our OASIS-trained models on ADNI (and vice-versa), we move beyond simple AUC reporting toward verifying the clinical stability of the learned brain signatures.

This rigorous validation approach is critical for assessing the feasibility of deploying these models in diverse real-world clinical settings.