New AI Reads Brain Scans Four Ways to Spot Alzheimer’s Stages and Know When It Is Unsure

Alzheimer’s disease is notoriously difficult to detect in its earliest stages, and the brain scans that clinicians rely on often hide their secrets in plain sight. A new study published in Discover Artificial Intelligence introduces a deep learning framework called MS-AtroNet that classifies individual brain MRI slices into four categories of cognitive severity: NonDemented, VeryMild, Mild, and Moderate. The work, led by S. K. Rakib Ul Islam Rahat of Wright State University together with an international team of collaborators, tackles a problem that has long frustrated researchers in computational neuroimaging: the visual differences between adjacent severity categories can be so subtle, spatially diffuse, and entangled with class imbalance that conventional image classifiers struggle to separate them reliably.

The technical heart of MS-AtroNet is its refusal to rely on a single way of seeing. The architecture combines three complementary representation streams. The first is a custom four-stage convolutional neural network backbone built from Multi-Receptive Field Attention Blocks, or MRFABs. Each MRFAB contains three parallel branches: a local texture branch using standard 3 by 3 convolutions, a directional anisotropic branch that applies 1 by 5 and 5 by 1 convolutions to emphasize horizontally and vertically oriented patterns, and a dilated contextual branch using a 3 by 3 convolution with dilation 3 to reach a 7 by 7 effective receptive field without a larger kernel. The outputs of these branches are concatenated, passed through squeeze-and-excitation channel attention with a reduction ratio of 8, and combined with a residual pathway before GELU activation.

The second stream brings in long-range contextual reasoning. Rather than processing the raw image, a VMamba-S module receives only the deepest Stage 4 feature map from the CNN backbone, a compact 7 by 7 grid with 512 channels. This grid is reshaped into 49 spatial tokens and processed by two Visual State-Space blocks that perform two-dimensional selective state-space scanning in four directions: left-to-right, right-to-left, top-to-bottom, and bottom-to-top. The outputs from these four routes are summed, gated by a SiLU-activated branch, and projected back to the original 512-dimensional shape. This design lets the model capture broad spatial dependencies that may relate to diffuse atrophy patterns, something a purely convolutional network with a limited effective receptive field cannot easily do.

The third stream is perhaps the most unusual: a frequency-domain pathway built on a one-level Haar discrete wavelet transform. The input MRI slice is decomposed into four subbands, LL, LH, HL, and HH, which respectively capture low-frequency structure, directional intensity transitions, and high-frequency texture and edge information. Each subband is processed through its own convolutional branch, and a small gating network assigns adaptive weights to the four subbands based on global average pooling descriptors. The weighted features are concatenated into a 128-channel wavelet representation, from which a single-channel frequency gate is derived. Crucially, this wavelet stream never makes its own prediction; its role is to modulate how the spatial and contextual streams are fused.

That fusion happens in a module called Cross-Stream Attention Fusion, or CSAF. The original CNN Stage 4 features serve as the queries, while the VMamba-enhanced features provide the keys and values, with eight attention heads operating over the 49 spatial positions. The resulting cross-attended representation is multiplied element-wise by the wavelet-derived frequency gate, allowing frequency information to shape which spatial locations matter most. A residual connection preserves the original deep features, and a position-wise feed-forward network with pre-normalization completes the module. A Hierarchical Feature Pyramid then integrates the fused Stage 4 representation with the intermediate Stage 2 and Stage 3 features, concatenating 128, 256, and 512 channels into an 896-dimensional vector that feeds the final classifier.

That final classifier is where the framework earns the word evidential in its name. Instead of a standard softmax head, MS-AtroNet uses an Evidential Classification Head that estimates class evidence and Dirichlet concentration parameters, producing both class probabilities and an image-level uncertainty score. The model is trained with an evidential deep learning objective that combines a negative log-likelihood term with a Kullback-Leibler regularization term, whose contribution is annealed linearly over the first 50 epochs. This formulation means the network does not merely say which category it prefers; it also expresses how confident it is, a property the authors argue is essential when the Moderate class contains only 64 test images out of nearly 4,000.

The performance numbers are striking. On an internal test set of 3,917 original images from the Kaggle Augmented Alzheimer’s MRI Dataset, MS-AtroNet achieved an accuracy of 0.978, a Macro-AUC of 0.993, an AUC-PR of 0.957, a Macro-F1 of 0.961, and macro sensitivity and specificity of 0.957 and 0.990, all with 38.4 million trainable parameters. It outperformed six carefully matched baselines, including ResNet-50, VGG-16, EfficientNet-B4, DenseNet-121, ViT-B/16, and a CNN with CBAM attention, the strongest of which reached a Macro-F1 of only 0.897. Calibration results were equally impressive, with an expected calibration error of 0.018, a maximum calibration error of 0.061, and a Brier score of 0.041, far better than any baseline. Under controlled synthetic perturbations such as Gaussian noise, brightness, contrast, and blur, the model showed the smallest mean Macro-F1 degradation of any system tested, at just 0.014.

Ablation experiments, validated with paired Wilcoxon signed-rank tests across ten random seeds and Holm correction, confirmed that each major component earns its place. Removing both the VMamba-S and wavelet streams caused the largest drop, a median Macro-F1 reduction of 0.041, while removing either stream alone cost 0.020 and 0.013 respectively. Replacing the MRFAB blocks with standard 3 by 3 convolutions cost 0.019, and replacing cross-stream attention fusion or the hierarchical pyramid also produced significant degradation. The uncertainty mechanism proved useful in a different way: when thresholds derived solely from the validation split were used to abstain on uncertain predictions, retained-case accuracy rose from 0.978 to 0.998 at the most restrictive operating point, with Macro-F1 climbing to 0.993 as coverage fell to roughly 80 percent. An exploratory external evaluation on 756 Kaggle ImageOASIS images yielded an accuracy of 0.903 and Macro-F1 of 0.867, showing partial transfer while also revealing sensitivity to dataset shift.

The authors are unusually candid about the limits of these results. Patient, examination, and volume identifiers were unavailable in the public dataset, so patient-level independence across the training, validation, and test splits could not be verified, and the reported figures must be read strictly as image-level performance under possible within-subject and within-volume dependence. The Moderate class, with only 550 training and 64 test images, showed the highest proportional error rate at 9.38 percent, and a blinded qualitative review of 36 of the 85 total errors attributed most mistakes, 72.2 percent, to ambiguity between adjacent severity categories rather than to obvious model failures. The Grad-CAM and saliency visualizations included in the study are explicitly described as qualitative post hoc views that were not validated by expert readers or anatomical annotations. The team emphasizes that patient-disjoint, multisite, prospective, and clinically validated evaluation remains necessary before any of this technology could approach the clinic. Still, as a demonstration of how convolutional spatial analysis, state-space contextual modeling, wavelet frequency gating, and evidential uncertainty can be woven into a single coherent architecture, MS-AtroNet offers a compelling template for the next generation of trustworthy medical image classifiers, and a reminder that knowing when an AI is unsure may matter as much as knowing what it predicts.

Subject of Research: Deep learning classification of Alzheimer's disease severity from brain MRI slices using multi-stream contextual, frequency-aware, and uncertainty-aware modeling

Article Title: Multi-stream contextual and frequency-aware deep learning with evidential uncertainty for Alzheimer’s MRI slice classification

Article References: Rahat, S. K. R. U. I., Islam, M., Islam, M. A., Alam, M. F., Hossain, M. J., Manik, M. M. T. G., Sakib, S. H., Rahman, M. M., & Wata, C. (2026). Multi-stream contextual and frequency-aware deep learning with evidential uncertainty for Alzheimer’s MRI slice classification. Discover Artificial Intelligence, 6(1), Article 1375. https://doi.org/10.1007/s44163-026-02337-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02337-2

Keywords: Alzheimer's disease, MRI classification, deep learning, MS-AtroNet, VMamba, wavelet transform, evidential deep learning, uncertainty estimation, medical image analysis, class imbalance, neurodegenerative disease, artificial intelligence