Classification of direct microscopic fungi images using optimized graph networks

Fungal infections pose a significant threat to immunocompromised patients due to their variable antifungal susceptibility and the limitations of conventional identification methods. Current approaches relying on microscopy, histopathology, and culture demand specialized training, often leading to delays and increased computational costs. To overcome these limitations, we introduce a deep learning approach for identifying uncultured fungi patches. Our model leverages optimized graph-based neural networks to classify images of direct microscopic examination, addressing the challenges of low data availability and lengthy identification processes. To further enhance the model’s performance, its hyperparameters are optimized using various swarm-based optimization methods. Our model can potentially eliminate the need for expert mycologists and lengthy biochemical identification processes. By employing few-shot learning and hyperparameter optimization, the model achieves remarkable accuracy in identifying direct, uncultured fungal images, reducing identification time from 10-14 days to hours. To the best of our knowledge, our approach achieves the highest accuracy on the DeFungi (93.1 %) and DIFaS (100 %) datasets, surpassing the closest benchmark methods, which report accuracies of 86.8 % on DeFungi and 93.9 % on DIFaS.