Hierarchical probabilistic tissue modelling with deep learning for Alzheimer's disease detection from fluid-attenuated inversion recovery magnetic resonance imaging
Engineering Applications of Artificial Intelligence, vol.181, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 181
- Publication Date: 2026
- Doi Number: 10.1016/j.engappai.2026.115508
- Journal Name: Engineering Applications of Artificial Intelligence
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Compendex, INSPEC, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO)
- Keywords: Artificial intelligence for Alzheimer's disease detection, Deep learning, Explainable neuroimaging analysis, Fluid-attenuated inversion recovery magnetic resonance imaging, Hierarchical Gaussian mixture modelling, Three-dimensional convolutional neural network
- Azerbaijan State University of Economics (UNEC) Affiliated: Yes
Abstract
Detecting Alzheimer's disease (AD) from fluid-attenuated inversion recovery magnetic resonance imaging (FLAIR MRI) places two demands on artificial intelligence systems that the literature often treats separately: clinical readability of the model output and faithfulness to the imaging protocol. We address both within a single pipeline that pairs a hierarchical recursive mixture model (RMM) with a fixed-depth deep learning classifier. In the present work, the term recursive refers to two-stage probabilistic refinement to dynamic depth, conditional computation, or the Mixture-of-Recursions Transformer family. Sagittal three-dimensional (3D) FLAIR Digital Imaging and Communications in Medicine (DICOM) series are converted to Neuroimaging Informatics Technology Initiative (NIfTI) volumes, registered to Montreal Neurological Institute (MNI) space, and decomposed by a two-stage Gaussian mixture model, coarse cerebrospinal fluid, gray matter (GM), and white matter (WM) split, followed by tissue-specific refinement of GM and WM into finer sub-components. The resulting six probabilistic channels feed a 3D convolutional neural network (CNN) whose spatial attention is steered by the GM posterior. The system is evaluated on 240 unique subjects (120 AD and 120 cognitively normal) using a strict subject-level split with no participant appearing in more than one partition. On the held-out set the model reached 77.08% accuracy, 91.67% sensitivity, 62.50% specificity, 70.97% precision, an F-measure 1 score of 0.800, and a receiver operating characteristic area under the curve (ROC-AUC) of 0.7708. The classification-only RMM + 3D CNN variant scored higher (79.17% accuracy, ROC-AUC 0.875) than the full multi-task segmentation-classification model (75.0% accuracy, ROC-AUC 0.847).