Explainable convolutional neural network architectures for high-performance taxonomic classification of gasteroid...

Area of Science:

• Mycology
• Computational Biology
• Artificial Intelligence

Background:

• Gasteroid fungi are morphologically diverse and taxonomically challenging due to convergent evolution and closed fruiting bodies.
• Accurate classification is crucial for understanding fungal biodiversity and ecological roles.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for the accurate classification of six gasteroid macrofungi species.
• To assess the performance, efficiency, and interpretability of various convolutional neural networks (CNNs) for fungal identification.

Main Methods:

• A dataset of 1200 high-resolution images of six macrofungi species was used.
• Eleven pre-trained CNNs (DenseNet121, ResNeXt, RepVGG, ShuffleNetV2) were fine-tuned for classification.
• Explainable AI techniques (Grad-CAM, Guided Backpropagation) were employed for model interpretability.

Main Results:

• DenseNet121 achieved the highest accuracy (96.11%), F1-score (96.09%), and AUC (99.89%).
• ShuffleNetV2 demonstrated the fastest inference time (0.80 s), while RepVGG showed the highest energy efficiency (16.5%).
• Explainable AI methods highlighted biologically relevant image regions, enhancing model transparency.

Conclusions:

• Deep learning models can effectively classify gasteroid fungi with high accuracy and transparency.
• The proposed framework is scalable and adaptable for broader applications in biological classification and biodiversity monitoring.
• This AI-driven approach offers a robust solution for automated biodiversity assessments.