Residual neural networks in single instance-driven identification of fungal pathogens

Abstract

The rise in fungal infections, attributed to various factors including medical interventions and compromised immune systems, necessitates rapid and accurate identification methods. While traditional mycological diagnostics are time-consuming, machine learning offers a promising alternative. Nevertheless, the scarcity of well-curated datasets is a significant obstacle. To address this, a novel approach for identifying fungi in microscopic images using Residual Neural Networks and a subimage retrieval mechanism is proposed, with the final step involving the implementation of majority voting. The new method, applied to the Digital Images of Fungus Species database, surpassed the original patch-based classification using Convolutional Neural Networks, obtaining an overall classification accuracy of 94.7% compared to 82.4% with AlexNet FV SVM. The observed MCC metric exceeds 0.9, while AUC is near to one. This improvement is attributed to the optimization of hyperparameters and top layer architecture, as well as the effectiveness of the Mish activation function in ResNet-based architectures. Noteworthy, the proposed method achieved 100% accurate classification for images from 8 out of 9 classes after majority voting and is high resistant to overfitting, highlighting its potential for rapid and accurate fungal species identification in medical diagnostics and research.