Enhancing Virus Classification with Deep Learning: A Novel Approach using TEM Images

Virus Classification Study Overview

This study presents a systematic approach to virus classification using Transmission Electron Microscopy (TEM) images. The workflow consists of three main steps: preprocessing of TEM images, feature extraction through convolutional models, and final classification using a multi-layer perceptron (MLP).

Preprocessing of TEM Images

The researchers utilized two standard Digital Image Processing (DIP) methods. One method focused on filtering images, while the other involved domain transformation. These processed images were then used as inputs for separate convolutional models to extract relevant features.

Dataset Description

The dataset utilized is called the TEM virus dataset, available on Mendeley Data. This dataset, published in 2021, contains microscopic images of various viruses captured using two electron microscopes: LEO (Zeiss, Germany) and Tecnai 10. The images were taken after treating samples with phosphate-buffered saline and staining with phosphotungstic acid.

To standardize the dataset, all images were resized to 1 nm pixel size. Researchers faced several challenges, such as limited annotations, noise, and varying image resolutions. These issues were addressed by cropping images into patches and augmenting them to balance sample sizes. The dataset contains grayscale images in TIFF format and represents 14 different viruses.

Image Preparation for the Convolutional Branches

First Convolutional Branch:

1. Image Resizing: Original images were resized to 128 x 128 pixels.
2. Filtering: A local standard deviation filter was applied. The filter works by moving a window across the image to calculate local properties.

Second Convolutional Branch:

1. Transformation: A 2D Discrete Cosine Transform (DCT) was applied to the images to extract important features useful for classification.

Classification Model Architecture

The classification model comprises two convolutional branches and an MLP. Each branch uses several convolutional layers with different activation functions. The principal steps are:

1. Convolutional Layers: These layers extract features from the input images.
2. Max-Pooling Layers: They reduce the spatial dimensions of the data.
3. Batch Normalization: This step normalizes the inputs to improve training speed and stability.
4. Dropout Layers: Used to prevent overfitting.
5. Flattening: This process converts the extracted feature maps into vectors for the MLP.

The MLP consists of dense layers that process the combined features and classify the input samples based on the highest probability score assigned to each class.

Conclusion

This structured approach simplifies dataset handling and enhances virus classification accuracy using advanced image processing techniques. The integration of convolutional models and MLP yields a comprehensive classification process, demonstrating the effectiveness of the proposed methods in distinguishing between different types of viruses.