2026 ECNP Congress

EP02-2002

Neuroimaging-based artificial intelligence models for differential diagnosis between Alzheimer’s disease and frontotemporal dementia: a comparative evaluation

Cazacu Tudor — A.B. Ciubară, A. Ciubară, I. Chiscop1,, Ș.L.B. Burlea · "Dunărea de Jos" University of Galați, Medicine, Galati, Romania

2026 ECNP Congress

Introduction

Distinguishing Alzheimer’s disease from frontotemporal dementia has always been clinically demanding. Recently it has become demanding in a different way. With lecanemab approved for early Alzheimer’s disease, misclassification carries direct therapeutic implications, not only prognostic ones. A patient with behavioural variant frontotemporal dementia who is diagnosed as having Alzheimer’s and treated with an anti-amyloid antibody receives a drug that does not work and may cause harm. AI-based neuroimaging analysis has been proposed as a way to improve accuracy and reproducibility of this distinction, and the technical performance reported in research settings is genuinely impressive. Whether that performance survives contact with clinical reality is largely unanswered.

Aims

To compare neuroimaging-based AI approaches for the Alzheimer’s versus frontotemporal dementia differential, across structural MRI, FDG-PET, amyloid and tau positron emission tomography, deep learning architectures, multimodal fusion, and explainability methods, with explicit attention to what clinical translation requires beyond benchmark performance.

Methods

Systematic narrative review (2018–2025) using ADNI, the NINDS frontotemporal dementia dataset (NIFD), the Genetic Frontotemporal dementia Initiative (GENFI), and the ARTFL-LEFFTDS Longitudinal cohort. Architectures evaluated: 3D CNNs, vision transformers, Swin transformers, graph attention networks, and multimodal fusion models. Explainability methods: Grad-CAM, SHAP, and saliency maps. Studies were identified through PubMed and major neuroimaging conference proceedings.

Results

Three-dimensional CNNs reach AUC 0.89 for binary Alzheimer’s versus behavioural variant frontotemporal classification and 0.84 for multi-class subtype discrimination on benchmark data. Graph attention networks applied to cortical thickness distributions achieve state-of-the-art multi-class performance, capturing distributed atrophy patterns that purely local convolutional operations miss. FDG-PET automated analysis reaches 85–91% accuracy, and attention-based models generate metabolic pattern maps aligned with known clinical syndrome signatures. Amyloid PET offers categorical separation of Alzheimer’s from primary frontotemporal dementia in most cases, although cost and access remain prohibitive in routine practice.

The generalisation problem is more serious than the benchmark figures suggest. Cross-cohort performance is systematically lower than within-cohort performance across every architecture tested. NIFD’s roughly 300 cases — about one tenth the scale of ADNI — produce a class imbalance that biases classification toward Alzheimer’s disease. Scanner harmonisation using ComBat or neuroHarmonize is necessary but has limited effect on small frontotemporal cohorts. Explainability analysis confirms that models highlight anatomically plausible regions, the hippocampus for Alzheimer’s and the frontal cortex for behavioural variant frontotemporal dementia, but anatomical plausibility does not guarantee cross-cohort validity, as Bron and colleagues showed. What clinicians actually need from these tools — calibrated probabilistic output, reliable performance on atypical cases, seamless PACS integration — has not been comprehensively evaluated in any published model.

Conclusions

The technical performance of neuroimaging-based AI for the Alzheimer’s versus frontotemporal dementia differential is genuinely promising. The gap between benchmark performance and clinical readiness, however, remains wide. Closing it will require prospective multi-site validation under real clinical imaging conditions, expansion of frontotemporal data consortia, and health economic assessment of multimodal acquisition pathways.

References

References

  1. Chételat G. et al., 2020. Amyloid-PET and 18F-FDG-PET in the diagnostic investigation of Alzheimer’s disease and other dementias. Lancet Neurol. 19(11), 951–962.
  2. Nguyen M. et al., 2022. Deep grading for MRI-based differential diagnosis of Alzheimer’s disease and frontotemporal dementia. Artif. Intell. Med. PMC10904714.
  3. Ossenkoppele R. et al., 2018. Discriminative accuracy of [18F]flortaucipir PET for Alzheimer disease vs other neurodegenerative disorders. JAMA 320(11), 1151–1162.
  4. Selvaraju RR. et al., 2017. Grad-CAM: visual explanations from deep networks via gradient-based localisation. Proc. ICCV 618–626.
  5. Wen J. et al., 2020. Convolutional neural networks for classification of Alzheimer’s disease: overview and reproducibility. Med. Image Anal. 63, 101694.