Enhancing Microscopic Fungi Classification through Advanced Pretraining in Deep Learning Models

Fungl infections are a growing global health concern, with increasing prevalence, delayed diagnoses, and a mortality rate higher than previously estimated. Traditional diagnostic methods, such as culture-based and serological techniques, are time-consuming and require specialized expertise, limiting their broad applicability. Recent advancements in deep learning (DL) offer a promising alternative by enabling automated and accurate fungal image classification using microscopic images. However, existing DL approaches often rely on models pre-trained on ImageNet-1K, leading to domain shifts and suboptimal performance for medical imaging tasks. To overcome these limitations, we employ state-of-the-art architectures, including ConvNeXt and Swin Transformer, pretrained on ImageNet-22K to improve performance. Our results show that ConvNeXt-B achieves a 1.5% improvement in accuracy, while Swin T-B demonstrates a 0.8% increase over their ImageNet-1K pretrained counterparts, emphasizing the benefits of larger and more diverse pretraining datasets. Moreover, ConvNeXt-B outperforms established baseline models such as ResNet-101 and MobileNetV3, highlighting the superiority of advanced architectures in fungi image classification.