TEM virus images: Benchmark dataset and deep learning classification

• We publish a new challenging dataset with 1245 TEM images of 22 virus classes. propose baseline classification for this dataset. Our best model, fine-tuned DenseNet201, achieved 93.1% accuracy on the test set. custom CNN 90.1% despite being 10x smaller than DenseNet201. show importance application knowledge in design and interpretation. To achieve full potential deep learning (DL) models, such as understanding interplay between model (size), training strategy, amount data, researchers developers need access to dedicated image datasets; i.e., annotated collections representing real-world problems all their variations, complexity, limitations, noise. Here, we present, describe make freely available an transmission electron microscopy (TEM) It constitutes interesting challenge many practical applications virology epidemiology; e.g., detection, segmentation, classification, novelty detection. also present benchmarking results detection recognition using some top-performing (large small) networks well handcrafted very small network. compare evaluate transfer from scratch hypothesizing that limited dataset, is crucial good performance large network whereas our performs relatively when scratch. This one step towards how much data needed given task. The benchmark contains representative split into training, validation, sets Moreover, different established DL solution classifying subset 14 most-represented classes DenseNet201 pre-trained ImageNet set, 0.921 F1-score proposed Public real biomedical datasets are important contribution necessity increase shortcomings, requirements, improvements solutions or deploying clinical settings. compared hypothesize limited-sized achieving models. Last but not least, demonstrate creating models analyzing results.