Abstract
Alzheimers Dement. 2025 Dec;21 Suppl 8:e110254. doi: 10.1002/alz70862_110254.
ABSTRACT
BACKGROUND: Convolutional neural networks (CNNs) excel in classifying dementia subtypes from T1-weighted MRI scans, yet their limited interpretability hinders clinical adoption. Classical machine learning (ML) models, while typically less performant, offer greater explainability by utilising predefined anatomically-informed features. However, a systematic comparison of the many popular T1 MRI preprocessing pipelines for feature extraction - particularly in the context of dementia - is lacking.
METHOD: This study evaluates seven widely used MRI preprocessing pipelines (FSL, SPM, CAT12 VBM, CAT12 SBM, SynthSeg, Freesurfer, FastSurfer) across five open-source dementia imaging datasets (ADNI, OASIS, AIBLE, NIFD, NACC). Regional volumetric and surface-based features extracted by each pipeline were input into six classical ML classifiers: support vector machines (SVM), random forest, logistic regression, naïve Bayes, k-nearest neighbors (kNN), and XGBoost. Performance metrics, including accuracy, precision, sensitivity, and AUC-ROC, were used to evaluate the classification of dementia subtypes (including Alzheimer's disease and frontotemporal dementia). Additionally, raw MRI scans are classified using several 3D CNN architectures, including a modified ResNet18, to benchmark feature-based ML models against CNN performance. The analysis pipeline is summarised in Figure 1.
RESULT: Preliminary analysis on the ADNI dataset identified Freesurfer as the best-performing preprocessing pipeline, achieving a mean accuracy of 0.872 across ML classifiers. Random forest emerged as the top-performing model overall, with a mean accuracy of 0.840 across pipelines. Shapley value analysis for Freesurfer features revealed consistent influential regions across ML models; for example, Figure 2 highlights similarities between SVM and logistic regression models. The CNN results are currently being generated and will be available for comparison at the conference.
CONCLUSION: This study highlights the strengths and limitations of different preprocessing-ML combinations for dementia classification, providing a framework for optimising performance while enhancing interpretability. By systematically comparing these pipelines to CNNs, we aim to identify clinically viable alternatives that balance accuracy, explainability, and computational efficiency.
PMID:41433757 | DOI:10.1002/alz70862_110254
UK DRI Authors
