Virus-FusionNet: an explainable deep learning-based approach for virus classification from TEM images using DenseNet and lightweight hierarchical semi-local feature extraction network

Abstract Viruses cause a wide range of diseases, from influenza to AIDS, representing a major global health concern. Traditional methods for virus identification, such as Transmission Electron Microscopy (TEM), are time-consuming and labour-intensive. TEM image analysis is highly dependent on expert interpretation, which can delay the diagnosis and treatment. To address these limitations, this research proposes a novel deep learning model for automated virus classification using TEM images. The model utilises a dual-track architecture combining DenseNet for local feature extraction and a customised Lightweight Hierarchical Semi-Local Feature Extraction Network (LHSLFN) for capturing global and semi-local features. The Shuffled Flexi-Fusion Enhanced Convolutional (SFFEC) block in the LHSLFN track aims at extracting global features, while the Dynamic Multi-path Semi-Local Enhancement (DMSLE) block enhances local and semi-local feature extraction. The Multidimensional Collaborative Attention (MCA) mechanism refines the concatenated features, enhancing the focus of the model on crucial regions. The Virus-FusionNet was evaluated on the TEM virus dataset and achieved an accuracy of 99.73%, outperforming existing works and state-of-the-art networks. Additionally, the Virus-FusionNet provides interpretability through Grad-CAM and SHAP visualisations, facilitating an understanding of its decision-making process.