Responsible artificial intelligence in medical imaging: a systematic review

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Health Informatics

Background:

• Responsible artificial intelligence (AI) in medical imaging demands high diagnostic accuracy, transparent reasoning, equitable performance, privacy protection, calibrated uncertainty, and clinical trustworthiness.
• The integration of AI in medical diagnostics is rapidly advancing, necessitating a thorough evaluation of its responsible implementation.

Purpose of the Study:

• To systematically review and synthesize evidence on responsible AI in medical imaging.
• To identify current trends and gaps in the evaluation of AI for disease detection and diagnostic support across various imaging modalities.

Main Methods:

• A PRISMA-informed systematic review of 24 studies published between 2020 and 2025.
• Searches conducted across major scientific databases (PubMed, Scopus, Web of Science, etc.).
• Qualitative appraisal of extracted evidence using adapted QUADAS-2 and PROBAST-AI domains.

Main Results:

• Explainability methods (e.g., Grad-CAM, LIME, SHAP) were dominant, while fairness, privacy-preserving learning, and uncertainty estimation were less represented.
• Studies covered diverse applications including lung diseases, cancers, and diabetic retinopathy across X-ray, CT, MRI, and other modalities.
• High reported accuracies (e.g., >90%) should be interpreted cautiously due to reliance on internal validation, curated datasets, and limited demographic reporting.

Conclusions:

• Responsible medical imaging AI requires multidimensional evaluation beyond diagnostic accuracy.
• Future AI development must prioritize external validation, subgroup analysis, calibration, privacy assessment, explainability, and post-deployment monitoring for clinical trustworthiness.