A Multimodal CNN for Dementia Screening: Integrating Structural MRI and Clinical Data for Radiologist Support
Vatsal Mitesh Tailor1, Avery Buehler2, Tamanna Dhore2, Laya Krishnan2, SungJoon Won2, Vijaya Kolachalama3, Suguna Pappu2, Aimee Yu-Ballard2, Sridhar Seshadri1
1University of Illinois Urbana-Champaign, 2Carle Illinois College of Medicine, 3Boston University
Objective:
To develop and deploy a multimodal convolutional neural network (CNN) model that classifies dementia using structural T1 MRI scans and clinical features. This model aims to provide rapid, objective screening to augment radiologist workflow and reduce diagnostic delays.
Background:
Dementia diagnosis is often delayed, with missed cases common in primary care. Radiologists struggle with subtle, variable atrophy patterns on MRI. Successful clinical adoption requires interpretable AI that integrates into the clinical workflow, reduces interpretation time, and provides attention maps to boost confidence in screening.
Design/Methods:
We developed a multimodal framework incorporating a 3D CNN for MRI analysis and integrated clinical features (demographics, medical history, neuropsychology). Preprocessing included skull stripping, normalization, and MNI152 registration. Trained on ADNI and NACC data totaling 1,314 participants (aged 49-96), the model performs 3-way classification: cognitively normal (CN), mild cognitive impairment (MCI), and Alzheimer’s disease (AD). Explainability is provided via SHAP-generated MRI heatmaps and clinical importance scores.
Results:
The model achieved 84% accuracy on a 25-case radiologist-validated cohort. SHAP heatmaps successfully highlighted disease-specific patterns, including hippocampal atrophy and ventricular enlargement. AD cases showed concentrated attention in medial temporal and parietal regions, while CN cases exhibited minimal activation. The model processes a full scan in under two minutes on a standard GPU, enabling immediate decision support during clinical review.
Conclusions:
This multimodal AI framework integrates into the clinical workflow by providing objective screening and interpretable predictions, allowing radiologists to efficiently verify results. It is compatible with standard T1 MRI and clinical data; this compatibility supports scalable deployment for better outcomes in at-risk populations. Ongoing work is focused on validating the model on a larger, more diverse cohort to confirm its real-world performance.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.