DUNE

DUNE encodes multisequence brain MRI into low-dimensional embeddings to enable prediction of clinical and molecular phenotypes from full-size neuroimaging data.

Key Features:

• UNet-based autoencoder architecture: DUNE uses a UNet-based autoencoder trained on 3,814 selected scans including morphologically normal brains (healthy volunteers) and abnormal brains (glioma patients).
• Low-dimensional representations: The encoder generates low-dimensional embeddings from full-sized multisequence brain MRI scans to reduce data dimensionality for downstream analysis.
• Clinical and diagnostic predictive performance: Embeddings derived from healthy volunteer scans predict volumetry metrics, cardiovascular disease risk (AUROC=0.80), alcohol consumption (AUROC=0.99), and APOE4 allele status (AUROC=0.67).
• Generalization across disease cohorts: Embeddings were evaluated on independent Alzheimer’s disease, schizophrenia, and glioma cohorts totaling 1,322 individuals, achieving diagnosis prediction performance such as Alzheimer’s dementia (AUROC=0.92) and schizophrenia (AUROC=0.64).
• Glioma-specific predictions: For glioma scans, embeddings predict survival (C-index=0.608) and IDH molecular status (AUROC=0.92), with performance comparable to task-oriented models.
• Compatibility with basic machine learning algorithms: The low-dimensional embeddings are intended for use with basic machine learning algorithms for downstream predictive tasks.

Scientific Applications:

• Imaging-derived trait prediction: Use embeddings to predict volumetry metrics and imaging-linked risk factors such as cardiovascular disease risk, alcohol consumption, and APOE4 status.
• Disease classification: Apply embeddings for diagnostic classification in Alzheimer’s disease and schizophrenia, as demonstrated by reported AUROC values.
• Oncology prognostication and molecular prediction: Use embeddings for glioma patient prognostication (survival C-index) and molecular status prediction (IDH AUROC).
• Cross-cohort studies and translational research: Leverage transferable embeddings for analyses across multiple independent cohorts to support early diagnosis, personalized treatment planning, and study of neurodegenerative and neuropsychiatric disorders.

Methodology:

UNet-based autoencoder trained on 3,814 multisequence MRI scans produces low-dimensional embeddings from full-size scans, and those embeddings were used with basic machine learning algorithms and evaluated on independent cohorts totaling 1,322 individuals.