A Comparative Analysis of CNN Architectures, Fusion Strategies, and Explainable AI for Fine-Grained Macrofungi...

Area of Science:

• Mycology
• Computational Biology
• Artificial Intelligence

Background:

• Accurate identification of morphologically similar macrofungi is challenging for fungal taxonomy and biodiversity monitoring.
• Existing methods struggle with fine-grained visual classification in fungi.
• Automated identification tools are needed to support taxonomic and ecological studies.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automated classification of seven morphologically similar coprinoid macrofungi species.
• To compare the performance of various state-of-the-art convolutional neural networks (CNNs) and novel fusion models.
• To utilize Explainable AI (XAI) techniques to understand model decision-making processes.

Main Methods:

• A curated dataset of 1692 high-resolution macrofungi images was utilized.
• Ten state-of-the-art CNNs and three novel fusion models were evaluated.
• Performance metrics included accuracy, Matthews Correlation Coefficient (MCC), and Area Under the Curve (AUC).
• Explainable AI (XAI) methods like Grad-CAM and Integrated Gradients were employed.

Main Results:

• The Dual Path Network (DPN) achieved the highest accuracy (89.35%) as a single model.
• Feature-level fusion of Xception and DPN demonstrated competitive performance (88.89% accuracy).
• Lighter models (LCNet, MixNet) exhibited lower accuracy (72.05%).
• XAI confirmed models focused on key morphological features like caps and gills.

Conclusions:

• Deep learning models, especially deeper architectures and fusion models, are effective for fine-grained fungal classification.
• The developed framework offers a robust and interpretable tool for automated fungal identification.
• This approach has significant implications for advancing mycology, biodiversity research, and taxonomy.