Peering Inside the Black Box: Explainable AI to Interpret Advanced Computer Vision Fungal Pathogen Prediction

Antimicrobial resistance is a growing concern, with pathogenic fungi making a substantial contribution to untreatable life-threatening infections across the globe. Artificial intelligence is increasingly used in microbiology and antimicrobial resistance research, with promise to improve clinical diagnostics and infectious disease treatments. However, how AI models make predictions remains largely unknown, which is a major hurdle for human trust and regulatory approval. We train vision transformers (Swin Transformer-Tiny and Vision Transformer-Base 16) and convolutional neural networks (DenseNet121 and InceptionV3) to quickly and accurately identify pathogenic yeast species from microscopy images. Using explainable AI (Occlusion Sensitivity and Grad-CAM), we identify biologically relevant features (organelle, cell interior, cell wall, budding patterns/scars, and optical patterns) and irrelevant image features (background artifacts) these high-performance computer vision models may be using for pathogenic yeast classification. These findings improve AI-based microscopy classification of pathogenic yeast species and advance our understanding of the visual features these models rely on to make predictions.