Image annotation and curation in radiology: an overview for machine learning practitioners

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Science

Background:

• The efficacy of machine learning (ML) and artificial intelligence (AI) models heavily relies on the quality of training data.
• Inconsistent, non-standardized, or improperly de-identified data can lead to unreliable AI algorithms, particularly in sensitive fields like radiology.

Purpose of the Study:

• To provide a comprehensive overview of techniques and freely available software solutions for ensuring high-quality data in machine learning applications within radiology.
• To guide researchers in preparing datasets that are consistent, standardized, traceable, correctly annotated, and de-identified in compliance with data protection regulations.

Main Methods:

• Review of essential medical imaging concepts (resolution, pixel depth) and data storage formats.
• Exploration of free software for image processing (e.g., ImageJ, 3D Slicer) and anonymization techniques.
• Discussion of patient privacy protection methods, including anonymization and pseudonymization, adhering to GDPR and HIPAA.
• Overview of image annotation tools and data harmonization/normalization techniques.

Main Results:

• Identified key considerations for data quality in AI, including standardization, traceability, and de-identification.
• Highlighted the utility of free software solutions for medical image processing and annotation.
• Emphasized the importance of regulatory compliance (GDPR, HIPAA) in data preparation.
• Demonstrated methods for removing identifying features from radiological images.

Conclusions:

• Implementing robust data curation strategies is vital for developing reliable and compliant AI in radiology.
• Leveraging freely available software tools can significantly facilitate the preparation of high-quality datasets for AI research.
• Adherence to data privacy regulations is paramount throughout the dataset preparation lifecycle.