<p>Early and accurate identification of malaria parasites in high-resolution microscopic blood smear images is critical for timely clinical intervention and effective disease control, especially in resource-constrained healthcare environments. This study presents a novel hybrid classification framework, Vision Transformer with Kolmogorov–Arnold Network-based Mixture of Experts (ViT-KANMoE), that strategically integrates self-attention-based global feature extraction, functional decomposition via learnable B-spline activations, and adaptive expert learning to enhance diagnostic precision. The Vision Transformer (ViT) component captures global spatial dependencies and contextual features from blood smear images, effectively encoding variations in cell morphology and parasite distribution. The Kolmogorov–Arnold Network (KAN) module is implemented as a <i>Pure KAN</i> formulation, in which all nonlinear transformations are performed exclusively through learnable cubic B-spline basis functions along network edges, with no fixed activation functions such as ReLU, GELU, or Mish. This design grounds the nonlinear transformations directly in the Kolmogorov–Arnold representation theorem, providing adaptive feature refinement through learnable B-spline edge functions. Subsequently, a Mixture of Experts (MoE) layer employs multiple specialized expert subnetworks, each a Pure KAN, that process the ViT features in parallel and map them directly to class logits, with outputs combined through uniform aggregation. A key finding of this work is that uniform expert aggregation, where all experts contribute equally, consistently outperforms attention-based gating (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(99.26\%\)</EquationSource> </InlineEquation> vs. <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(97.78\%\)</EquationSource> </InlineEquation>), suggesting that complementary expert contributions are more effective than selective routing for this classification task. Extensive experiments on the malaria blood smear dataset from Central South University, Changsha, China demonstrate that ViT-KANMoE achieves a classification accuracy of <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(99.26\% \)</EquationSource> </InlineEquation> across seven clinically relevant cell categories, with consistently high per-class precision and recall. These findings establish ViT-KANMoE as a promising architecture for reliable and automated malaria diagnosis in clinical settings.</p>

ViT-KANMoE: a vision transformer enhanced with a pure Kolmogorov–Arnold network-based mixture-of-experts for malaria blood smear image classification

期刊论文
发表日期：

2026年6月25日

查看全文

Aluri Jitendra Chowdary,
Mukesh Medikonda,
Jasti Ramakrishna,
Jyostna Devi Bodapati

摘要

Early and accurate identification of malaria parasites in high-resolution microscopic blood smear images is critical for timely clinical intervention and effective disease control, especially in resource-constrained healthcare environments. This study presents a novel hybrid classification framework, Vision Transformer with Kolmogorov–Arnold Network-based Mixture of Experts (ViT-KANMoE), that strategically integrates self-attention-based global feature extraction, functional decomposition via learnable B-spline activations, and adaptive expert learning to enhance diagnostic precision. The Vision Transformer (ViT) component captures global spatial dependencies and contextual features from blood smear images, effectively encoding variations in cell morphology and parasite distribution. The Kolmogorov–Arnold Network (KAN) module is implemented as a Pure KAN formulation, in which all nonlinear transformations are performed exclusively through learnable cubic B-spline basis functions along network edges, with no fixed activation functions such as ReLU, GELU, or Mish. This design grounds the nonlinear transformations directly in the Kolmogorov–Arnold representation theorem, providing adaptive feature refinement through learnable B-spline edge functions. Subsequently, a Mixture of Experts (MoE) layer employs multiple specialized expert subnetworks, each a Pure KAN, that process the ViT features in parallel and map them directly to class logits, with outputs combined through uniform aggregation. A key finding of this work is that uniform expert aggregation, where all experts contribute equally, consistently outperforms attention-based gating ( \(99.26\%\) vs. \(97.78\%\) ), suggesting that complementary expert contributions are more effective than selective routing for this classification task. Extensive experiments on the malaria blood smear dataset from Central South University, Changsha, China demonstrate that ViT-KANMoE achieves a classification accuracy of \(99.26\% \) across seven clinically relevant cell categories, with consistently high per-class precision and recall. These findings establish ViT-KANMoE as a promising architecture for reliable and automated malaria diagnosis in clinical settings.