A comprehensive and easy-to-use multi-domain multi-task medical imaging meta-dataset

Area of Science:

• Medical Image Analysis
• Machine Learning
• Artificial Intelligence in Healthcare

Background:

• Machine learning (ML) integration revolutionizes medical image analysis.
• A key challenge is the scarcity of large, diverse, and well-annotated medical datasets.
• Medical images require extensive preprocessing due to variations in format and size.

Purpose of the Study:

• Introduce the Medical Imaging Meta-Dataset (MedIMeta) to address data limitations.
• Provide a standardized, multi-domain, multi-task dataset for ML in medical imaging.
• Facilitate easier adoption and development of ML models for medical image analysis.

Main Methods:

• Developed MedIMeta, a novel meta-dataset comprising 19 individual datasets.
• Covered 10 distinct medical imaging domains and 54 different medical tasks.
• Standardized all data to a uniform format compatible with ML frameworks like PyTorch.

Main Results:

• MedIMeta offers a readily usable resource for ML research.
• Technical validation demonstrated utility in fully supervised learning.
• Cross-domain few-shot learning baselines showed promising results.

Conclusions:

• MedIMeta effectively tackles the challenge of data scarcity in medical ML.
• The standardized, comprehensive nature of MedIMeta promotes broader ML application in healthcare.
• This resource can accelerate advancements in AI-driven medical diagnostics and analysis.