Lightweight Dual-Branch AI Reads Brain Scans to Spot Alzheimer’s Stages

A deep learning framework that pairs a convolutional neural network with a lightweight Swin Transformer has achieved 98.42 percent accuracy in classifying Alzheimer’s disease stages from structural MRI scans, according to a study published in Frontiers in Computational Neuroscience. The framework, named NeuroCST, was designed to capture both the fine-grained local details of brain atrophy and the broader spatial context of the image at once, while keeping the model small enough to run on modest hardware.

The motivation behind the work is a practical bottleneck in computer-aided diagnosis. Deep learning models using structural MRI have shown strong performance in detecting Alzheimer’s disease, but the rising number of model parameters and high computational demands make deployment difficult in resource-limited clinical environments. Many prior approaches rely on computationally expensive three-dimensional architectures, and the authors also note that reliance on large-scale pretraining on natural image datasets does not consistently transfer well to medical imaging because of the substantial domain shift between the two.

NeuroCST addresses this with a parallel dual-branch design. The first branch is an enhanced EfficientNet-based convolutional network that extracts fine-grained morphological and textural features using depthwise separable convolution and residual learning. Depthwise separable convolution splits the work into a depthwise operation, where a single filter acts on each feature map, and a pointwise 1×1 convolution that exchanges information across maps, cutting both computational cost and parameter counts substantially compared with dense convolution.

The second branch is a customized lightweight Swin Transformer, dubbed SwinLite, which models global contextual dependencies. Standard vision transformers compute self-attention between all pairs of image tokens, an operation whose cost scales quadratically with image size. SwinLite instead uses separable self-attention, which reformulates the matrix-based attention operation as element-wise products and reduces complexity to scale linearly with image dimensions. The branch also employs shifted-window partitioning, which lets attention windows connect to neighboring windows so that distant dependencies are not lost, and it shrinks the transformer’s MLP expansion factor from four to one to save computation.

After the two branches process an image independently, their feature maps are concatenated, and a new spatial attention module called NeuroCST-SA refines the merged representation. The module computes spatial attention weights through convolutional layers followed by a sigmoid function, producing a map that emphasizes regions with diagnostic significance, such as hippocampal, entorhinal cortex, and temporal lobe atrophy, while suppressing redundant responses. The authors report that this post-fusion refinement helps discriminate subtle inter-class contrast and texture variations shared among different dementia stages.

The framework was evaluated on a publicly available Kaggle dataset, AD-Multiclass Images DatasetV2, derived from the clinically annotated OASIS MRI cohort. The dataset contains 6,400 original and 33,984 augmented MRI images, 40,384 in total, spread across four classes: non-demented, very mild demented, mild demented, and moderate demented. The scans were converted from 3D NIfTI volumes into 2D JPEG slices of 224 by 224 pixels. The study performed secondary analysis only, with no new participant recruitment or MRI acquisition.

The experiments used two setups. In the first, the 33,984 augmented images served for training and validation while the 6,400 original images were reserved for testing. In the second, only the 6,400 original images were used, split into 4,097 training, 1,023 validation, and 1,280 testing images, with augmentation applied only to the training subset. On the Case 1 original test set, NeuroCST-SA reached 98.42 percent accuracy, 98.50 percent sensitivity, 98.60 percent precision, and an F1-score of 98.55 percent, with a standard deviation of 2.49 percent across holdout cross-validation. In the Case 2 setup using original data only, accuracy reached 98.67 percent.

The comparisons were extensive. NeuroCST-SA outperformed conventional CNN architectures tested under identical training conditions, including InceptionResNetV2 at 86.95 percent, GoogleNet at 87.48 percent, InceptionV3 at 88.59 percent, DarkNet-53 at 88.13 percent, SqueezeNet at 89.77 percent, VGG-16 at 91.25 percent, and DarkNet-19 at 93.53 percent. It also exceeded vision transformer models, including Tiny ViT at 89.14 percent, Cross ViT at 92.97 percent, LocalViT at 93.08 percent, Dual-ViT at 94.52 percent, and the hybrid LeViT at 96.78 percent. An ablation study showed the components building on each other: the customized CNN branch alone reached 97.31 percent accuracy, SwinLite alone 97.74 percent, the combined NeuroCST 98.18 percent, and NeuroCST-SA 98.42 percent. The authors caution, however, that these stepwise differences of 0.24 to 0.87 percentage points are smaller than the reported standard deviations, so they should be read as trends rather than statistically established gains.

Efficiency is where the design most clearly pays off. The base NeuroCST model has 15.3 million parameters and requires 5.2 GFLOPs, while the NeuroCST-SA variant is slightly smaller at 14.8 million parameters, 4.9 GFLOPs, and an inference time of 11.4 milliseconds per image. By comparison, the tested InceptionResNetV2 carried roughly 55 million parameters and 15.2 GFLOPs, and VGG-16 about 138 million parameters. Replacing the standard Swin Transformer with SwinLite alone cut parameters from 12.4 million to 9.6 million, about 23 percent fewer, and FLOPs from 4.6 billion to 3.8 billion, about 17 percent fewer, without a drop in accuracy. The authors suggest this balance makes the framework a candidate for clinical settings with limited computing resources.

The study does carry limitations the authors acknowledge. NeuroCST works on individual 2D slices without inter-slice volumetric context, and many slices may contain little disease-relevant information; future work will explore ROI-guided 2.5D and 3D analysis. Evaluation relied on a single Kaggle dataset derived from OASIS without independent multicenter validation, so cross-site and patient-level generalization remains to be established. Augmentation may also introduce distributional dependence, and the absence of subject identifiers prevented verification of subject-independent splitting. The authors also note that no 3D model was trained in the study, so no quantitative comparison with volumetric approaches was made. As a classification research tool rather than a deployed diagnostic system, the framework’s reported figures describe performance on this dataset, and the authors frame it as a promising base for future clinically deployable automated Alzheimer’s classification, with possible extensions incorporating non-image data such as age, gender, and neuropsychological scores.

Subject of Research: A lightweight dual-branch CNN and Swin Transformer deep learning framework for classifying Alzheimer's disease stages from structural MRI scans.

Article Title: Efficient neuroanalytic local–global learning with dual-branch CNN–lightweight Swin Transformer for Alzheimer’s disease analysis

Article References: Khan, S. H., Iqbal, R., Alqahtani, S., Shan, R. U., Alhefdhi, A., Alkhamees, B., Gasmi, K., & Ali, A. (2026). Efficient neuroanalytic local–global learning with dual-branch CNN–lightweight Swin Transformer for Alzheimer’s disease analysis. Frontiers in Computational Neuroscience, 20, 1972923. https://doi.org/10.3389/fncom.2026.1972923

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1972923

Keywords: Alzheimer's disease, deep learning, structural MRI, convolutional neural network, Swin Transformer, computer-aided diagnosis, spatial attention, lightweight neural networks, medical imaging, dementia classification, OASIS dataset, neuroscience