A Novel ConvNeXtV2-MIL Approach for Accurate and Efficient Classification of Microscopic Fungal Morphology

Abstract

Fungal infections, especially in people with weakened immune systems, are a significant global health burden. Accurate identification of fungal morphology from microscopic images is a critical step in guiding timely antifungal treatment decisions. However, manual morphological assessment remains highly dependent on expert mycologists and is prone to inter-observer variability. In this study, we propose a hybrid deep learning framework that integrates the ConvNeXtV2-Base architecture with a Multi-Head Attention-based Multiple Instance Learning module for automated classification of microscopic fungal morphology images. The framework was evaluated on the open-access DeFungi dataset, consisting of 3696 microscopic images representing five clinically meaningful fungal morphology classes. In comparative experiments, classical vision transformer (ViT) models achieved 91.20% accuracy, while MIL-enhanced ViT models reached 93.99%. The proposed ConvNeXtV2-Base + MIL hybrid method outperformed all evaluated architectures, achieving 98.90% classification accuracy. These results establish a new benchmark for automated fungal morphology classification and highlight the potential of AI-assisted decision-support tools to aid expert mycologists in morphology-based assessment workflows.