Multimodal PET/MRI radiomics model for the classification of Alzheimer's disease and other dementia subtypes: A multicenter study

Accurate and early diagnosis of Alzheimer's disease (AD) remains a major clinical challenge. This study aims to explore the value of multimodal positron emission tomography (PET) and magnetic resonance imaging (MRI) radiomics to improve diagnostic accuracy.

To develop and evaluate hippocampus-based multimodal PET/MRI radiomics models for AD classification and exploratory five-class classification, and to compare early fusion (EF) with multi-criteria decision-making (MCDM)-based decision-level fusion.

A total of 758 patients who underwent both PET ([18F]FDG and Amyloid-β) and MRI (3DT1, T1WI, T2WI, T2-FLAIR) were retrospectively included as the primary dataset in this study. 945 radiomic features (RFs) per sequence were extracted from the bilateral hippocampus. Models were evaluated on the primary cohort using a five-fold stratified cross-validation at the subject level (80% training and 20% testing in each fold). Feature selection was performed independently for each sequence using the Least Absolute Shrinkage and Selection Operator (LASSO). For multimodal PET/MR models, features from all six sequences were concatenated, and feature selection was applied (EF). Six classifiers were used to construct classification models for two tasks, including (1) AD/non-AD classification, and (2) exploratory five-class classification among AD, FTD, VaD, MCI, and other dementia. In addition, an MCDM-based strategy was implemented to fuse multimodal RFs and multi-classifiers. Modality-matched external validation was performed in an independent cohort of 45 patients with 3DT1 and Aβ PET available. Decision curve analysis (DCA) was employed to assess the clinical usefulness of the models. SHapley Additive exPlanations (SHAP) was applied to explore the interpretability of the EF-based model.

Among the 758 patients in the primary dataset, AD, FTD, VaD, and MCI were present in 481 (63.5%), 34 (4.5%), 15 (2.0%), 136 (17.9%) cases, with 92 (12.1%) classified as the other dementia subtypes. The MCDM-based models achieved the highest area under the receiver operating characteristic curves (AUCs) for both tasks on the primary dataset (Task 1: AUC = 0.848; Task 2: AUC = 0.797), significantly higher than EF-based models in Task 2 (AUC = 0.782, p < 0.001) and showing comparable performance in Task 1 (AUC = 0.839, p = 0.629). On the external dataset, EF and MCDM models achieved AUCs of 0.741 and 0.761, respectively, for Task 1. The MCDM model showed a higher AUC than EF. DCA suggested potential net-benefit advantages for the multimodal models across selected threshold ranges. SHAP analysis identified wavelet-transformed texture features as the key predictors.

Hippocampus-based PET/MRI radiomics models showed promising performance for binary AD classification, with further improvement achieved by the MCDM strategy. The five-class findings should be regarded as exploratory. External validation provides preliminary support for the reduced 3DT1-Aβ PET model and does not establish the generalizability of the complete six-modality framework.