Madhurie Kumar Seth, K. Srinivas, A. Charan Kumari
{"title":"Classifying fungi biodiversity using hybrid transformer models","authors":"Madhurie Kumar Seth, K. Srinivas, A. Charan Kumari","doi":"10.1016/j.mimet.2025.107155","DOIUrl":null,"url":null,"abstract":"<div><div>Fungi are essential members of ecosystems, playing key roles in nutrient cycling, agriculture, and medicine. Their classification into proper species helps us to understand their biodiversity, allowing us to leverage their ecological and practical benefits. A new hybrid deep learning-based technique has been proposed, merging the Vision Transformer and Swin Transformer models with transfer learning frameworks like MobileNetV2, DenseNet121, and EfficientNetB0 for Fungi multiclass classification. This study utilized a publicly available dataset containing 9115 images of five fungal species from UC Irvine Machine Learning Repository. To address significant class imbalance, several data augmentation techniques were employed. The results showed that the Swin Transformer combined with DenseNet121 achieved the highest classification accuracy of 96.96 % for training, 95.97 % for validation, and 95.57 % for testing, while other models like ViT-DenseNet121 and Swin-MobileNetV2 also delivered competitive results. Using confusion matrices and benchmark classification metrics, and paired statistical testing, the analysis highlights the models' ability to generalize effectively and minimize misclassifications. To further ensure the robustness of the findings, a five-fold cross-validation was performed across all hybrid models. Additionally, explainable AI techniques, specifically Grad-CAM visualizations, were employed to interpret the model's focus areas, confirming attention to biologically significant structures. This research demonstrates a balance between modeling local features and capturing global context. Indeed, these hybrid models prove to be scalable and efficient for complex biological datasets. This interdisciplinary study bridges ecology and advanced technology by applying deep learning to enhance fungal classification. This study aims to improve the management and understanding of fungal biodiversity for the promotion of conservational and sustainable practices for the betterment of our ecosystem. The findings have significant applications, including sustainable agriculture through early detection of fungal plant pathogens, improved medical diagnostics for fungal infections, and biodiversity conservation through precise species monitoring.</div></div>","PeriodicalId":16409,"journal":{"name":"Journal of microbiological methods","volume":"236 ","pages":"Article 107155"},"PeriodicalIF":1.7000,"publicationDate":"2025-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of microbiological methods","FirstCategoryId":"99","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0167701225000715","RegionNum":4,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q4","JCRName":"BIOCHEMICAL RESEARCH METHODS","Score":null,"Total":0}
引用次数: 0
Abstract
Fungi are essential members of ecosystems, playing key roles in nutrient cycling, agriculture, and medicine. Their classification into proper species helps us to understand their biodiversity, allowing us to leverage their ecological and practical benefits. A new hybrid deep learning-based technique has been proposed, merging the Vision Transformer and Swin Transformer models with transfer learning frameworks like MobileNetV2, DenseNet121, and EfficientNetB0 for Fungi multiclass classification. This study utilized a publicly available dataset containing 9115 images of five fungal species from UC Irvine Machine Learning Repository. To address significant class imbalance, several data augmentation techniques were employed. The results showed that the Swin Transformer combined with DenseNet121 achieved the highest classification accuracy of 96.96 % for training, 95.97 % for validation, and 95.57 % for testing, while other models like ViT-DenseNet121 and Swin-MobileNetV2 also delivered competitive results. Using confusion matrices and benchmark classification metrics, and paired statistical testing, the analysis highlights the models' ability to generalize effectively and minimize misclassifications. To further ensure the robustness of the findings, a five-fold cross-validation was performed across all hybrid models. Additionally, explainable AI techniques, specifically Grad-CAM visualizations, were employed to interpret the model's focus areas, confirming attention to biologically significant structures. This research demonstrates a balance between modeling local features and capturing global context. Indeed, these hybrid models prove to be scalable and efficient for complex biological datasets. This interdisciplinary study bridges ecology and advanced technology by applying deep learning to enhance fungal classification. This study aims to improve the management and understanding of fungal biodiversity for the promotion of conservational and sustainable practices for the betterment of our ecosystem. The findings have significant applications, including sustainable agriculture through early detection of fungal plant pathogens, improved medical diagnostics for fungal infections, and biodiversity conservation through precise species monitoring.
期刊介绍:
The Journal of Microbiological Methods publishes scholarly and original articles, notes and review articles. These articles must include novel and/or state-of-the-art methods, or significant improvements to existing methods. Novel and innovative applications of current methods that are validated and useful will also be published. JMM strives for scholarship, innovation and excellence. This demands scientific rigour, the best available methods and technologies, correctly replicated experiments/tests, the inclusion of proper controls, calibrations, and the correct statistical analysis. The presentation of the data must support the interpretation of the method/approach.
All aspects of microbiology are covered, except virology. These include agricultural microbiology, applied and environmental microbiology, bioassays, bioinformatics, biotechnology, biochemical microbiology, clinical microbiology, diagnostics, food monitoring and quality control microbiology, microbial genetics and genomics, geomicrobiology, microbiome methods regardless of habitat, high through-put sequencing methods and analysis, microbial pathogenesis and host responses, metabolomics, metagenomics, metaproteomics, microbial ecology and diversity, microbial physiology, microbial ultra-structure, microscopic and imaging methods, molecular microbiology, mycology, novel mathematical microbiology and modelling, parasitology, plant-microbe interactions, protein markers/profiles, proteomics, pyrosequencing, public health microbiology, radioisotopes applied to microbiology, robotics applied to microbiological methods,rumen microbiology, microbiological methods for space missions and extreme environments, sampling methods and samplers, soil and sediment microbiology, transcriptomics, veterinary microbiology, sero-diagnostics and typing/identification.