[Paper Review] Monkeypox Image Data collection

Md Manjurul Ahsan, Muhammad Ramiz Uddin+1 more|arXiv (Cornell University)|Jun 3, 2022

Poxvirus research and outbreaks32 citations

TL;DR

The paper presents a public Monkeypox image dataset assembled from web sources, expanded to about 1905 images via augmentation, to enable ML-based diagnosis research.

ABSTRACT

This paper explains the initial Monkeypox Open image data collection procedure. It was created by assembling images collected from websites, newspapers, and online portals and currently contains around 1905 images after data augmentation.

Motivation & Objective

• Motivate creation of a public Monkeypox image dataset for computational analysis due to lack of available data.
• Describe the data collection process from websites, newspapers, and online portals.
• Show data augmentation to increase dataset size and diversity for machine learning.
• Highlight potential applications of the dataset in training image-based diagnostic models and mobile tools.

Proposed method

• Assemble Monkeypox, Chickenpox, Measles, and Normal skin images from public sources.
• Apply data augmentation using Keras ImageDataGenerator to expand the dataset.
• Provide sample images and histogram analyses to illustrate image characteristics.
• Make the dataset publicly available via a GitHub repository.
• Reference related datasets and prior COVID-19 image data collection efforts to justify approach.

Experimental results

Research questions

• RQ1 Can a public Monkeypox image dataset be effectively constructed from open web sources for computational analysis?
• RQ2 How does data augmentation affect dataset size and variability for Monkeypox-related image data?
• RQ3 Is the resulting dataset suitable for training image-based diagnostic models (e.g., CNNs, GANs, transfer learning) for Monkeypox?
• RQ4 What are the visual and statistical characteristics (e.g., histograms) of the collected images compared to other skin diseases?

Key findings

• The dataset includes 43 Monkeypox, 47 Chickenpox, 17 Measles, and 54 Normal original images.
• Augmentation increased the dataset to approximately 1905 images.
• The dataset supports exploring image-based Monkeypox diagnosis and could reduce reliance on PCR or microscopy.
• Sample figures and histograms illustrate image characteristics and distributions.
• The dataset is publicly available on GitHub for community use.