MedMNIST v2 - A large-scale lightweight benchmark for 2D and 3D biomedical image classification

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Standardized biomedical datasets are crucial for developing and validating machine learning models.
• Existing datasets often lack standardization or require specialized domain knowledge.
• A need exists for a comprehensive, accessible collection of biomedical images for diverse research applications.

Purpose of the Study:

• Introduce MedMNIST v2, a large-scale, standardized dataset of 2D and 3D biomedical images.
• Provide a versatile resource for classification tasks across various dataset scales and complexities.
• Facilitate research and education in biomedical image analysis, computer vision, and machine learning.

Main Methods:

• Compiled 18 datasets (12 2D, 6 3D) of biomedical images.
• Pre-processed all images to a uniform size (28x28 for 2D, 28x28x28 for 3D).
• Included corresponding classification labels for each image, ensuring no background knowledge is needed.

Main Results:

• MedMNIST v2 comprises 708,069 2D images and 9,998 3D images.
• The dataset supports diverse tasks: binary/multi-class classification, ordinal regression, and multi-label classification.
• Benchmarking of baseline methods using 2D/3D neural networks and AutoML tools was performed.

Conclusions:

• MedMNIST v2 offers a valuable, accessible resource for advancing biomedical image analysis and machine learning.
• The standardized nature and diverse tasks make it suitable for a wide range of research and educational purposes.
• Public availability of data and code promotes collaboration and innovation in the field.