TEM virus images: Benchmark dataset and deep learning classification.

Area of Science:

• Biomedical image analysis
• Deep learning applications
• Virology and epidemiology

Background:

• Deep learning (DL) models require large, annotated image datasets for optimal performance.
• Existing datasets may not capture the complexity and noise of real-world biological imaging.
• A dedicated dataset is needed to understand the interplay of model size, training strategy, and data volume.

Purpose of the Study:

• To present and release a novel annotated transmission electron microscopy (TEM) image dataset.
• To benchmark deep learning models for virus detection, segmentation, and classification.
• To investigate the effectiveness of transfer learning versus training from scratch on limited data.

Main Methods:

• A benchmark dataset comprising 1245 TEM images across 22 virus classes was curated.
• A representative split for training, validation, and testing was established.
• Performance of various deep learning networks, including large and small models, was evaluated for virus classification.

Main Results:

• The best model, DenseNet201 (pre-trained on ImageNet and fine-tuned), achieved 0.921 F1-score and 93.1% accuracy.
• Transfer learning significantly improved performance for large models on the limited dataset.
• A handcrafted small network demonstrated competitive performance when trained from scratch.

Conclusions:

• Publicly available biomedical datasets are essential for advancing deep learning in healthcare.
• Transfer learning is critical for large deep learning models when dataset size is limited.
• Domain expertise is vital for creating effective datasets and interpreting deep learning model results.