Checklist for Reproducibility of Deep Learning in Medical Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Reproducible Research

Background:

• Deep learning (DL) offers transformative potential in medical prognosis, diagnosis, and treatment planning.
• Transparent documentation is crucial for enhancing reproducibility and refining DL techniques in medicine.
• Existing reporting guidelines may not fully address the unique challenges of DL in medical imaging.

Purpose of the Study:

• To develop a comprehensive and validated checklist for reporting deep learning models in medical imaging.
• To enhance the reproducibility and reliability of DL applications in the medical field.
• To provide a standardized framework for documenting DL models used in healthcare.

Main Methods:

• A modified Delphi method involving 11 experts in medical imaging and DL was employed.
• A preliminary checklist was compiled from existing literature and guidelines.
• Two survey rounds using Likert scales and content validity ratio (CVR) were conducted to refine items and achieve consensus.

Main Results:

• The Delphi process resulted in a final 26-item checklist for reporting DL models.
• The checklist demonstrated high face and content validity, with a CVR cutoff of 0.59.
• The final checklist was refined through expert consensus, excluding non-essential items and incorporating new suggestions.

Conclusions:

• The developed 26-item checklist facilitates reproducible reporting of DL tools in medical imaging.
• This standardized checklist empowers scientists to replicate study results and enhances the reliability of DL applications.
• The checklist addresses the need for transparent documentation in the rapidly evolving field of medical DL.