Uncertainty quantification for artificial intelligence in medical imaging: what every radiologist needs to know

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Artificial intelligence (AI) is increasingly used in clinical radiology, requiring radiologists to interpret AI outputs and oversee AI systems.
• Radiologists need a deeper understanding of technical AI concepts like uncertainty quantification (UQ) for safe deployment.
• UQ estimates the reliability of AI predictions, crucial for identifying when AI outputs require cautious interpretation.

Purpose of the Study:

• To introduce key uncertainty quantification (UQ) concepts relevant to clinical radiology.
• To differentiate between aleatoric and epistemic uncertainty in AI models.
• To review UQ methods used in AI imaging research and practice.

Main Methods:

• A narrative review of recent AI imaging studies.
• Summarization of commonly used UQ approaches.
• Illustration of UQ application in practice through selected studies.

Main Results:

• UQ methods are being applied in AI imaging studies to assess prediction reliability.
• Identified trends, findings, and limitations in the practical application of UQ.
• Highlighted challenges including calibration, thresholding, computational cost, and need for validation.

Conclusions:

• Uncertainty quantification (UQ) can enhance the safety and interpretability of AI-assisted radiology screening.
• Further research and clinical validation are necessary to overcome current challenges in UQ implementation.
• Addressing UQ challenges is vital for the responsible integration of AI in radiology.