AI-Based Differential Diagnosis of Neurodegenerative Diseases Using Structural MRI: A Systematic Review
Abstract
Accurate differential diagnosis among neurodegenerative diseases is a major clinical challenge due to overlapping symptomatology and pathologies. Artificial intelligence (AI) applied to structural MRI (sMRI) offers a promising, noninvasive approach for identifying subtle, disease-specific patterns of brain atrophy. This systematic review evaluates the landscape of AI methods for this purpose, focusing on the differential diagnosis between neurodegenerative diseases while excluding isolated disease-versus-control comparisons. Following the PRISMA guidelines, we analyzed 82 studies published between January 2008 and December 2025. Our analysis reveals a distinct methodological evolution from classical machine learning (ML), which relies on handcrafted features, to end-to-end deep learning (DL) architectures that automatically learn representations directly from image data. While models often report high classification accuracy (typically exceeding 90% for binary tasks), accuracy ranges from 70% to 92% for standard 3-class tasks, and drops to 65% to 88% for broader multi-disease frameworks (≥ 4 classes) or out-of-domain evaluations. Furthermore, we identify systemic barriers that delay clinical translation: a widespread reliance on private datasets (77% of studies), a severe lack of rigorous out-of-domain validation (only 28% of studies), a low rate of public code availability (20% of studies), and limited use of pathologically or biologically confirmed diagnostic labels (only 18% of studies). These challenges in data infrastructure and validation standards undermine model generalizability and cast doubt on the clinical relevance of reported performance metrics. To bridge the gap between technical benchmarks and clinical utility, future research must prioritize collaborative data sharing, robust external validation, and improved model interpretability, ultimately translating AI tools into trustworthy diagnostic applications.
Related
- HMGB1 functional plasticity in neurodegenerative diseases and its translational implications
- Research landscape of neurodegenerative diseases and brain energy metabolism: A multi-database integrative analysis
- Diagnosing Neurodegenerative Diseases by Label-Free 3-D Imaging of Intestinal Samples
- Loss of TDP-43 function drives cryptic circular RNAs in neurodegenerative diseases
- Novel Tau Interaction Partners as Potential Biomarkers and Therapeutic Targets in Neurodegenerative Diseases
BibTeX
@article{navet2026aibased,
title = {{AI-Based Differential Diagnosis of Neurodegenerative Diseases Using Structural MRI: A Systematic Review}},
author = {Éloi Navet and Rémi Giraud and Boris Mansencal and Andrew Zamai and Vincent Planche and Pierrick Coupé},
journal = {Imaging Neuroscience},
year = {2026},
month = oct,
doi = {10.1162/imag.a.1448},
url = {https://hal.science/hal-05640352v2},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - AI-Based Differential Diagnosis of Neurodegenerative Diseases Using Structural MRI: A Systematic Review
AU - Éloi Navet
AU - Rémi Giraud
AU - Boris Mansencal
AU - Andrew Zamai
AU - Vincent Planche
AU - Pierrick Coupé
PY - 2026
DA - 2026/10/20
JO - Imaging Neuroscience
DO - 10.1162/imag.a.1448
UR - https://hal.science/hal-05640352v2
AB - Accurate differential diagnosis among neurodegenerative diseases is a major clinical challenge due to overlapping symptomatology and pathologies. Artificial intelligence (AI) applied to structural MRI (sMRI) offers a promising, noninvasive approach for identifying subtle, disease-specific patterns of brain atrophy. This systematic review evaluates the landscape of AI methods for this purpose, focusing on the differential diagnosis between neurodegenerative diseases while excluding isolated disease-versus-control comparisons. Following the PRISMA guidelines, we analyzed 82 studies published between January 2008 and December 2025. Our analysis reveals a distinct methodological evolution from classical machine learning (ML), which relies on handcrafted features, to end-to-end deep learning (DL) architectures that automatically learn representations directly from image data. While models often report high classification accuracy (typically exceeding 90% for binary tasks), accuracy ranges from 70% to 92% for standard 3-class tasks, and drops to 65% to 88% for broader multi-disease frameworks (≥ 4 classes) or out-of-domain evaluations. Furthermore, we identify systemic barriers that delay clinical translation: a widespread reliance on private datasets (77% of studies), a severe lack of rigorous out-of-domain validation (only 28% of studies), a low rate of public code availability (20% of studies), and limited use of pathologically or biologically confirmed diagnostic labels (only 18% of studies). These challenges in data infrastructure and validation standards undermine model generalizability and cast doubt on the clinical relevance of reported performance metrics. To bridge the gap between technical benchmarks and clinical utility, future research must prioritize collaborative data sharing, robust external validation, and improved model interpretability, ultimately translating AI tools into trustworthy diagnostic applications.
SN - 2837-6056
ER -