MPF-KANSC: Multi-Plane sMRI AD Diagnosis

• The paper introduces MPF-KANSC, a deep-learning framework that fuses coronal, sagittal, and axial sMRI views to capture heterogeneous Alzheimer’s disease atrophy patterns.
• It employs a Kolmogorov-Arnold Network-guided spatial-channel attention module to flexibly model complex nonlinearities and improve feature discrimination.
• Empirical results on the ADNI dataset demonstrate enhanced classification accuracy, robust sensitivity/specificity, and anatomically interpretable insights compared to prior methods.

MPF-KANSC is an interpretable deep-learning framework for Alzheimer’s disease diagnosis and prediction of mild cognitive impairment conversion from structural MRI (sMRI). The name denotes Multi-Plane Fusion with Kolmogorov-Arnold Network-guided Spatial-Channel Attention. It was proposed to address two limitations identified in prior sMRI-based models: many methods use only a single anatomical plane, and conventional attention modules rely on fixed nonlinearities such as ReLU or sigmoid, which can limit discrimination of complex and nonlinear pathological patterns. MPF-KANSC therefore combines parallel feature extraction from the coronal, sagittal, and axial planes with a KAN-guided attention mechanism intended to represent sMRI atrophy features more effectively and to improve interpretability through region localization and anatomical analysis (Yang et al., 8 Aug 2025).

1. Conceptual scope and problem setting

MPF-KANSC is designed for AD diagnosis and for distinguishing stable mild cognitive impairment (sMCI) from progressive mild cognitive impairment (pMCI) using sMRI (Yang et al., 8 Aug 2025). Its central premise is that neuroanatomical abnormalities associated with AD are spatially heterogeneous and may not be fully characterized from a single imaging plane. The framework therefore treats the three standard anatomical planes—axial, coronal, and sagittal—as complementary views rather than interchangeable projections.

The model also addresses a second representational problem: traditional attention modules commonly use static nonlinear mappings in channel recalibration. In MPF-KANSC, these mappings are replaced in the channel-attention pathway by a Kolmogorov-Arnold Network, which the paper describes as offering a more flexible nonlinear function approximation mechanism. This design is intended to facilitate more precise identification and localization of disease-related abnormalities (Yang et al., 8 Aug 2025).

Within the paper’s framing, interpretability is not auxiliary. The framework is explicitly presented as an interpretable architecture whose outputs are tied to top-activated regions, Grad-CAM visualization, AAL3 atlas mapping, and correlation analyses across disease stages. This suggests that the model is meant not only to classify but also to expose anatomically meaningful structure in the learned representation.

2. Network organization and data flow

The input is a 3D sMRI scan that is preprocessed and reoriented into the three standard anatomical planes: axial, coronal, and sagittal (Yang et al., 8 Aug 2025). MPF-KANSC then applies three parallel 3D CNN branches, one per plane, to extract plane-specific features. The branches share the same architecture but use independent parameters.

Each branch contains five hierarchical stages, beginning with an initial $2\times2\times2$ convolution and followed by several $3\times3\times3$ and $1\times1\times1$ convolutions, together with residual and concatenation blocks for multiscale feature reuse (Yang et al., 8 Aug 2025). This organization is intended to preserve both hierarchical abstraction and multiscale structural context. After branch-specific extraction, each plane’s features are passed through a KANSC attention module before cross-plane fusion.

The post-attention features from the coronal and sagittal branches are projected back to the axial space and then fused with axial features. In the operational description provided for the framework, fusion is performed via element-wise addition (Yang et al., 8 Aug 2025). The fused representation is then passed to a classification layer. The paper further states that this pipeline balances local and global feature learning through a hybrid loss composed of cross-entropy and smoothed logits-constraint terms.

A common simplification would be to describe the architecture as a conventional multi-branch CNN with attention. That would be incomplete. In the proposed formulation, the architectural novelty lies specifically in the conjunction of three-plane feature extraction, explicit inter-plane alignment, and KAN-based channel attention, rather than in any one component taken in isolation.

3. Multi-plane fusion as a structural representation strategy

The MPF component, or multi-plane fusion, is the framework’s mechanism for combining anatomical information from the coronal, sagittal, and axial planes (Yang et al., 8 Aug 2025). Each sMRI volume is processed into three separate 3D tensors, and each tensor is handled by its own branch. After attention weighting, the outputs are mapped to a common reference space—usually the axial plane—and fused.

The rationale stated in the paper is that important AD-related structural changes may be salient in one orientation but less prominent in another. By fusing features from all three planes, the model can capture more comprehensive structural information than single-plane approaches (Yang et al., 8 Aug 2025). The paper’s ablation study is reported to show that three-plane fusion consistently improves metrics over one-plane or two-plane configurations.

This multi-view formulation is particularly relevant for pathology characterized by subtle and distributed atrophic patterns. The paper connects the design to the difficulty of identifying complex and nonlinear relationships among pathological brain regions when only a single plane is used. A plausible implication is that MPF is meant to reduce orientation-specific blind spots in feature extraction rather than merely to increase input redundancy.

Operationally, MPF is not early fusion at the image level. It is feature-level fusion after plane-specific representation learning and attention enhancement. That distinction matters because it preserves plane-dependent specialization before integration. In the paper’s terminology, this supports richer and more discriminative representations of sMRI pathology (Yang et al., 8 Aug 2025).

4. KAN-guided spatial-channel attention

The KANSC module is the second defining component of MPF-KANSC. It consists of a spatial-attention submodule followed by a channel-attention submodule in which the standard multilayer perceptron is replaced by a Kolmogorov-Arnold Network (Yang et al., 8 Aug 2025).

For a feature map $F_k$ from plane $k$, spatial attention is computed from channel-wise maximum and average projections: $F_{k,\max}(d,h,w)=\max_c F_k(c,d,h,w),$

$$F_{k,\mathrm{avg}}(d,h,w)=\frac{1}{C}\sum_{c=1}^{C}F_k(c,d,h,w).$$

These two maps are concatenated along the channel dimension, passed through a 3D $3\times3\times3$ convolution and a sigmoid, producing a spatial attention map $M_{k,s}$. The spatially refined feature is then

$F'_k=M_{k,s}\odot F_k.$

Channel attention is subsequently computed by global average pooling: $3\times3\times3$0 followed by a KAN: $3\times3\times3$1 and channel reweighting: $3\times3\times3$2

The KAN itself is described as

$3\times3\times3$3

with each layer applying per-edge B-spline-parameterized univariate functions (Yang et al., 8 Aug 2025). The paper ties this construction to the Kolmogorov-Arnold Representation Theorem, stating that it provides a more flexible and accurate nonlinear function approximation technique than standard fixed nonlinearities.

Within the framework, KANSC is therefore not generic attention. Its stated purpose is twofold: spatially highlight salient voxels and channel-wise model complex nonlinear interdependencies among learned features. The paper further argues that this improves interpretability because the nonlinear transformation learned on each edge is more analyzable than conventional MLP-based channel attention (Yang et al., 8 Aug 2025).

5. Objective function and training formulation

The fused attention-weighted features are optimized with a hybrid loss: $3\times3\times3$4 with $3\times3\times3$5 (Yang et al., 8 Aug 2025). Here, $3\times3\times3$6 is the standard cross-entropy loss, while $3\times3\times3$7 is the smoothed logits-constraint loss.

According to the paper summary, the SLC term encourages the patch-level attention weights from the attention module to match the predicted class probabilities, so that feature selection and the final decision are more aligned (Yang et al., 8 Aug 2025). In the model’s stated interpretation, this improves optimization stability while balancing local and global feature learning. The role of the loss is therefore not only classification accuracy but also consistency between attention distributions and classifier outputs.

This training objective is tightly coupled to the model’s interpretability claims. If the attention mechanism is intended to support anatomical localization, then constraining attention patterns to track the predictive signal is a direct way to reduce the gap between what the network highlights and what it actually uses for classification. The paper presents this as part of the architecture rather than as a post hoc interpretability add-on.

6. Empirical results, interpretability, and reported findings

Experiments on the ADNI dataset are reported for the groups CN, sMCI, pMCI, and AD (Yang et al., 8 Aug 2025). On the CN vs AD task, MPF-KANSC achieved 94.3% accuracy, 93.1% sensitivity, 95.5% specificity, and 96.9% AUC. On the sMCI vs pMCI task, it achieved 84.2% accuracy, 83.5% sensitivity, 85.0% specificity, and 89.1% AUC (Yang et al., 8 Aug 2025). The paper states that the framework outperforms the compared methods, including 3D-ResAttNet, LA-GMF, and MSACNet.

The paper also reports that ablations confirm the contribution of both major components: multi-plane fusion and KANSC each incrementally improve performance, and three-plane fusion performs better than one-plane or two-plane alternatives (Yang et al., 8 Aug 2025). These results are presented as evidence that the gains are not attributable solely to a larger model or to generic attention.

Interpretability is developed through Grad-CAM visualization, AAL3 Atlas mapping, and correlation matrices that characterize how region-level attention patterns change from MCI to AD (Yang et al., 8 Aug 2025). The model is reported to consistently identify right-hemisphere subcortical regions—including Amygdala_R, Hippocampus_R, ParaHippocampal_R, Pallidum_R, and Thal_AV_R—as among the most discriminative regions in AD and pMCI versus sMCI. The paper further states that correlation matrices show weakening inter-regional coupling from pMCI to AD, with atrophy becoming more right-lateralized and less coordinated.

On that basis, the authors describe the framework as providing new evidence of right-lateralized asymmetry in subcortical structural changes during AD progression (Yang et al., 8 Aug 2025). This is framed as an interpretable biological finding rather than only a classification result. A cautious reading is that the model’s saliency and correlation analyses support the right-lateralization hypothesis within the experimental setting used, while the clinical generality of that pattern would require external validation beyond the reported study.

7. Position within sMRI-based AD modeling

Within the paper’s own comparison, MPF-KANSC is positioned against two recurrent tendencies in prior work: single-plane structural modeling and attention mechanisms built on fixed nonlinearities (Yang et al., 8 Aug 2025). Its contribution is to combine a multi-plane representation strategy with KAN-guided spatial-channel recalibration in a single framework.

The resulting system can be understood as a model of AD-related structure that is simultaneously multi-view, nonlinearity-adaptive, and interpretability-oriented. The paper does not present the method as a generic replacement for all neuroimaging classifiers; rather, it argues that this particular combination is advantageous for identifying subtle atrophic features in sMRI and for localizing disease-relevant abnormalities.

A common misconception would be to interpret the reported interpretability outputs as independent proof of neuropathology. The paper more carefully supports the claim that MPF-KANSC identifies and visualizes key regions and that its findings are consistent with prior evidence of right-lateralized subcortical change (Yang et al., 8 Aug 2025). Another misconception would be to reduce the method to “attention plus fusion.” In the paper’s formulation, the decisive design choice is that the channel-attention mechanism itself is reparameterized through a Kolmogorov-Arnold Network, thereby changing the nonlinear hypothesis class used to model inter-channel dependencies.

In summary, MPF-KANSC is an sMRI framework that integrates three-plane feature extraction, axial-space feature fusion, and KAN-guided spatial-channel attention under a hybrid classification loss. Its reported empirical value lies in improved AD-related classification on ADNI and in anatomically interpretable evidence of right-lateralized subcortical involvement during disease progression (Yang et al., 8 Aug 2025).