ViViMZheimer - a slice based end to end model for Alzheimer’s disease diagnosis from 3D MRI

Sep 29, 2026

Uploaded by Aditya

Slice-based models are widely used for Alzheimer’s disease (AD) identification because they reduce parameters and speed up inference, but existing methods still require extra slice extraction steps and often cannot operate end-to-end from MRI to diagnostic outputs. They also commonly rely on Transformer designs that introduce quadratic complexity when capturing inter-slice dependencies. ViViMZheimer directly processes 3D MRI, using a ViViT-inspired spatial encoder and Mamba-based temporal modeling with linear complexity, enhanced by spatial attention and a gated bottleneck convolution. Evaluated on ADNI, it achieves 98.17%, 82.21%, and 83.15% accuracy for AD vs CN, AD vs MCI, and CN vs MCI. The results support efficient automated AD diagnosis from 3D MRI.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages19

Reading Time48 min

AccessFree to download

Background

Alzheimer’s disease and cognitive stages

Deep learning for AD recognition

Related methods

Voxel-based, patch-based, and ROI-based approaches

Slice-based approaches and slice selection strategies

Technical challenge and proposal

Modeling inter-slice dependencies

ViViMZheimer: spatial encoder and temporal modeling

[www. nature.com/scientificreports](www. nature.com/scientificreports) OPEN ViViMZheimer a slice based end to end model for Alzheimer’s disease diagnosis from 3D MRI Juan Zhou1, Jiahui Wan1, Xia Chen2􀀍, Xiong Li1􀀍, Zejiu Wu3, Zhiwei Zhang1 & Chengjie Zhang1 Slice-based models have been widely applied in Alzheimer’s disease (AD) identification tasks due to their reduced parameter count and fast inference speed. However, existing slice-based models require additional slice extraction steps and cannot achieve an end-to-end process from MRI to diagnostic results. Moreover, they often rely on Transformer architectures to model inter-slice dependencies, which suffer from quadratic computational complexity. To address these limitations, we propose ViViMZheimer, a slice-based end-to-end model that directly processes 3D MRI data and generates diagnostic predictions. ViViMZheimer integrates aViViT-inspired spatial encoder with a Mambabased temporal modeling mechanism, maintaining linear computational complexity while effectively capturing inter-slice dependencies along three spatial orientations. Additionally, a lightweight spatial attention module emphasizes lesion-relevant brain regions, and a gated bottleneck convolution refines key features in later stages ofthe model. We evaluated ViViMZheimer on theADNI dataset, where it achieved accuracies of 98.17%, 82.21%, and 83.15% in distinguishing AD vs. cognitively normal (CN), AD vs. mild cognitive impairment (MCI), and CN vs. MCI, respectively. These results demonstrate that ViViMZheimer provides an effective and computationally efficient solution for automated Alzheimer’s disease diagnosis from 3D MRI scans. Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by extracellular plaques containing β-amyloid (Aβ) and intracellular neurofibrillary tangles composed of hyperphosphorylated tau. Clinically, AD most commonly presents with memory-related cognitive impairment, although it can also manifest with non-memory deficits such as impairments in expressive speech, visuospatial skills, and executive function. The clinical course and degree of cognitive impairment vary widely: many cases are sporadic rather than dominantly inherited, and some patients first report subjective cognitive decline even when objective tests remain within the normal range1. Mild cognitive impairment (MCI) denotes the symptomatic prodromal stage marked by mild deficits in one or more cognitive domains while overall functional abilities are largely preserved; individuals with MCI are at elevated risk for progression to AD, but may also remain stable or revert to cognitively normal (CN) status2–4. Accurate, objective, and efficient methods to discriminate AD, MCI, and CN subjects are therefore critical for early intervention and clinical management. Deep learning (DL), a subfield of machine learning built on deep, multi-layer neural architectures, excels at automatically learning hierarchical, nonlinear representations from raw data5. Applied to magnetic resonance imaging (MRI) and related neuroimaging modalities, DL methods can extract subtle, high-dimensional features that are difficult to capture with handcrafted descriptors, enabling more sensitive and specific stratification of Alzheimer’s disease. Recent work has leveraged these capabilities to push performance on AD classification tasks, but practical deployment still faces challenges related to model complexity, computational cost, and effective modeling of three-dimensional brain structure6. MRI-based DL approaches for AD recognition can be broadly grouped into four categories: voxel-based, patch-based, ROI-based, and slice-based models. Voxel-based methods process the entire 3D MRI volume to extract features such as intensity, shape, and texture. While they can achieve strong diagnostic performance for Alzheimer’s disease, they are computationally intensive duetothe large number of voxels, and may include regions o

FAQs

Why are slice-based models important for Alzheimer’s disease diagnosis from MRI?

They reduce parameter count and improve inference speed by converting volumetric MRI into 2D slices, which lowers input dimensionality and computational burden.

What limitations exist in prior slice-based approaches?

They typically require separate slice extraction steps and often lack true end-to-end processing from MRI to diagnostic predictions. They may also use Transformer-based mechanisms that increase computational complexity.

What are the main components of ViViMZheimer and how do they address inter-slice dependency modeling?

ViViMZheimer combines a ViViT-inspired spatial encoder with Mamba-based temporal modeling to capture inter-slice dependencies with linear computational complexity. It further uses a lightweight spatial attention module and a gated bottleneck convolution to refine lesion-relevant features.

Sep 29, 2026

Uploaded by Aditya