Interpretable mycology: leveraging Kolmogorov–Arnold networks for high-accuracy Lactarius species classification and comparative benchmarking

Abstract This study introduces a hybrid deep learning framework that integrates convolutional neural networks (CNNs) with Kolmogorov–Arnold networks (KAN) for fine-grained classification of eight Lactarius species using a curated dataset of 1.614 images. The proposed CNN–KAN architecture significantly outperforms seven state-of-the-art baseline models, ConvNeXt-Small, EfficientNetV2-Small, MobileNetV3-Small, RegNetY-400MF, ResNet-50, SqueezeNet1.1M, and ViT-Small, across all evaluation metrics. The model achieved an accuracy of 0.9877, F1-score 0.9877, precision 0.9879, sensitivity 0.9877, specificity 0.9982, MCC 0.9855, and AUC 0.9994, representing improvements of approximately 1–3% over high-capacity baselines such as ConvNeXt-Small (accuracy 0.9775), ResNet-50 (0.9754), and EfficientNetV2-Small (0.9672), and a substantial margin of + 23.36% compared with SqueezeNet1.1M (0.7541). Statistical analysis confirmed that CNN–KAN achieved the lowest Friedman mean rank (1.07) and demonstrated significant superiority over RegNetY-400MF, ViT-Small, and SqueezeNet1.1M in the Nemenyi post hoc test ( p < 0.05). Only three misclassifications occurred in 488 independent test samples, all within morphologically similar Lactarius aurantiacus images. Explainable AI analysis using LIME revealed that correct predictions predominantly relied on biologically meaningful structures, including gill lamellation, cap zonation, and stipe–cap transitions, while misclassifications were linked to background interference and chromatic ambiguity. Collectively, the findings demonstrate that the CNN–KAN framework provides a highly accurate, statistically validated, and interpretable solution for automated fungal taxonomy, with strong potential for deployment in ecological monitoring and digital biodiversity assessment pipelines. Moreover, the computational complexity associated with the hybrid CNN–KAN architecture, including high-dimensional feature extraction, spline-based functional transformations, and extensive hyperparameter optimization, necessitates the use of high-performance computing (HPC) resources. This highlights the model’s strong alignment with supercomputing frameworks for scalable, parallel, and real-time biodiversity analysis applications.