Revisiting medical image retrieval via knowledge consolidation

Area of Science:

• Artificial Intelligence in Medicine
• Digital Health Governance
• Medical Data Management

Background:

• AI and digital medicine require robust governance for ethical and secure healthcare implementation.
• Medical image retrieval is crucial for clinical decisions and patient data protection.
• Existing methods struggle with representative hash codes, OOD detection, and adversarial attacks.

Purpose of the Study:

• To develop an advanced AI method for medical image retrieval.
• To enhance the security and effectiveness of AI in healthcare systems.
• To address limitations in current hashing techniques and out-of-distribution challenges.

Main Methods:

• Proposed Depth-aware Representation Fusion (DaRF) to integrate shallow and deep features.
• Introduced Structure-aware Contrastive Hashing (SCH) using image fingerprints for adaptable pairings.
• Implemented content-guided ranking for improved retrieval robustness and reproducibility.

Main Results:

• The novel method effectively identifies out-of-distribution samples.
• Achieved significant improvements in medical image retrieval performance (p<0.05).
• Demonstrated a 5.6-38.9% increase in mean Average Precision on an anatomical radiology dataset.

Conclusions:

• The proposed AI method enhances medical image retrieval security and accuracy.
• DaRF and SCH contribute to a more secure AI-driven healthcare environment.
• The approach offers superior performance and robustness compared to existing methods.