The study, published in Frontiers in Artificial Intelligence, is built on a simple observation about the standard pipeline. Convolutional neural networks dominate Alzheimer’s classification from MRI, but the classifier head placed on top of the learned features is almost always a multilayer perceptron with fixed activation functions such as ReLU or GELU and linear weights. Such networks exhibit a spectral bias toward low-frequency functions and may need many parameters to represent the fine, highly non-linear texture and atrophy gradients that distinguish adjacent dementia stages. The researchers propose ConvNeXt–KAN, which retains a transfer-learned ConvNeXt-Tiny feature extractor but replaces the conventional head with a Kolmogorov–Arnold Network, an architecture in which the activations themselves are learned rather than fixed.

Kolmogorov–Arnold Networks are grounded in the Kolmogorov–Arnold representation theorem, a resolution of Hilbert’s thirteenth problem, which states that any multivariate continuous function can be written as a finite composition of continuous univariate functions and additions. Whereas a conventional multilayer perceptron applies a fixed non-linearity after a linear map, a KAN layer places a learnable univariate function on each edge and sums at the nodes, classically parameterized by B-splines. Evaluating and rescaling spline bases is computationally expensive, so the team adopted the FastKAN formulation, which proves that third-order B-splines can be closely approximated by Gaussian radial basis functions, yielding a substantially faster and numerically more stable implementation. Each KAN layer first applies layer normalization to keep activations within the effective support of the fixed grid, the main practical failure mode of radial-basis-function KANs, and the head stacks two such layers with dropout, reducing the 768-dimensional backbone features to a hidden width of 128 and then to four class logits.

The choice of backbone was deliberate. ConvNeXt-Tiny, with roughly 28 million parameters, offers a favorable capacity–regularization trade-off for a training set of only a few thousand images, while providing the modern design elements, including large depthwise kernels, inverted bottlenecks, layer normalization, and GELU activations, that make ConvNeXt competitive with Transformers. Training proceeded in two phases: the backbone was frozen while the randomly initialized KAN head adapted at a higher peak learning rate, then the whole network was fine-tuned at a much lower rate so the features could specialize in brain MRI without catastrophic forgetting. The learning rate followed a cosine schedule with linear warm-up, which matters because the randomly initialized head produces large early gradients that could destabilize the pre-trained backbone.

The training recipe was equally robustness-oriented. Each batch was regularized with MixUp or CutMix, schemes that form convex combinations or patch-wise pastes of image pairs and mix their labels in proportion, implementing a form of vicinal risk minimization that smooths the decision boundary between adjacent stages. Label smoothing discouraged overconfident logits and improved probability calibration, AdamW decoupled weight decay from the adaptive update to improve generalization, and test-time augmentation averaged predictions over the clean image and four lightly augmented copies, acting as a lightweight ensemble that reduces variance. At inference the model reached 94.5 percent accuracy, a macro-averaged F1 of 0.944, and a mean one-vs-rest ROC-AUC of 0.991, with moderate dementia classified essentially perfectly and mild dementia recognized with an F1 of 0.97. The remaining error concentrated on the clinically hardest boundary: non-demented recall was 0.82, with most misclassifications falling into the very-mild category, consistent with the near-normal appearance of very-mild disease.

Perhaps the most technically consequential part of the paper is its treatment of data. The OASIS collection used here is severely imbalanced: the moderate-dementia class contains only 488 axial slices against 67,222 non-demented slices, roughly two orders of magnitude fewer. A widespread remedy is to undersample every class to the smallest one, which discards roughly 98 percent of the data and starves precisely the under-represented, clinically difficult categories of training signal. Conversely, oversampling minority classes before splitting can place identical images in both training and test sets, leaking information and inflating accuracy. The researchers instead capped each abundant class at 1,500 slices, held out a fixed balanced number of images per class for validation and test, and balanced only the training partition by random replication, so no validation or test image was ever duplicated into training. Because augmentation is sampled afresh each epoch, replicated images appear in different forms rather than being merely memorized.

An ablation isolated the two main design choices. With a randomly initialized, frozen backbone the network collapsed toward a single class, reaching only 0.35 accuracy and 0.25 macro-F1. Replacing it with a compact CNN trainable from scratch on the minority-capped data recovered a usable model at 0.72 accuracy. Pre-trained ConvNeXt features together with the expanded-data protocol yielded the full model at 0.945 accuracy, confirming that pre-trained features and data utilization are the dominant factors. A head-to-head comparison under a matched shortened protocol showed that the KAN and MLP heads were statistically indistinguishable, with McNemar’s test returning p = 0.52, so the practical case for the KAN head rests on its learnable-activation capacity and negligible runtime cost rather than a clear accuracy gain in this regime. The KAN head added only 2.84 percent to total parameters, and per-image inference latency was actually marginally lower than for the MLP head.

The imbalance-handling comparison was more decisive. The proposed image-space resampling strategy outperformed focal loss, class-balanced cross-entropy, and feature-space SMOTE across accuracy, macro-F1, the Matthews correlation coefficient, and Cohen’s kappa, with the SMOTE and focal-loss differences statistically significant. The researchers attribute the wide SMOTE gap to the fact that interpolated feature vectors are only weakly constrained to lie on the manifold of real Alzheimer’s MRI features, so the classifier partially fits synthetic points that are not representative of clinical data. Post-hoc saliency maps computed with Grad-CAM and Score-CAM concentrated attention on the cortical rim and adjacent parenchymal regions rather than on background pixels or the skull outline, the pattern expected if the model relies on genuinely disease-relevant morphology rather than non-cerebral artifacts.

The external validation is where the study tempers its own headline. Evaluated without any fine-tuning on an independent four-class Alzheimer’s MRI collection distributed on Kaggle, acquired under different imaging conditions with different subjects, test-time-augmented accuracy dropped to 61.6 percent with a macro-averaged F1 of 0.565, a fall of roughly 33 percentage points in line with the degradation repeatedly observed when MRI classifiers are transported across acquisition sites without adaptation. The class-wise picture is more informative than the aggregate: moderate dementia continued to be recovered perfectly, replicating the internal behavior and consistent with the most visually distinctive MRI signature being least sensitive to domain shift, while mild-dementia recall collapsed from 98.63 percent to 10.96 percent and very-mild recall fell from 97.26 percent to 56.16 percent. The two clinically intermediate stages absorbed almost all of the additional error, precisely because the visual differences between adjacent stages are subtle even for expert raters, so the decision boundary in that narrow region is the one most sensitive to changes in image intensity statistics, contrast, and slice selection between datasets.

The authors also quantify the impact of the evaluation protocol itself. Many of the very high accuracies, often at or above 99 percent, reported on these data are obtained with image-level, per-slice train-test splitting, which can place near-duplicate slices of the same patient in both partitions, leaking subject identity and inflating measured performance relative to genuine patient-independent generalization. Although the protocol here removes image-duplication leakage by splitting before oversampling, the publicly distributed OASIS slices are not grouped by subject, so slices from one participant may still fall in different partitions, a limitation the team identifies as the most important next step alongside validation on the Alzheimer’s Disease Neuroimaging Initiative and other multi-site clinical cohorts. The gross ordering of class difficulty was preserved between the two datasets, which is non-trivial and suggests the learned features carry a genuine, if imperfect, signal about disease stage rather than a purely dataset-specific shortcut. Any clinical deployment, the study concludes, would require domain adaptation to the target site rather than direct reuse of OASIS-trained weights, and subject-level validation and multi-site external testing are essential before clinical claims.

Subject of Research: ConvNeXt–KAN: a Kolmogorov–Arnold Network head for robust multi-class staging of Alzheimer's disease from brain MRI

Article Title: ConvNeXt–KAN: a Kolmogorov–Arnold Network head for robust multi-class staging of Alzheimer's disease from brain MRI

Article References: ConvNeXt–KAN: a Kolmogorov–Arnold Network head for robust multi-class staging of Alzheimer's disease from brain MRI. (n.d.). https://doi.org/10.3389/frai.2026.1903281

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1903281

Keywords: Alzheimer's disease, Kolmogorov–Arnold Networks, brain MRI, deep learning, class imbalance, transfer learning, medical image classification, ConvNeXt, Kolmogorov, Arnold, Network, head