Examining explainable artificial intelligence in TNM staging with PET-CT: a user-centred observation study.

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Radiology and Nuclear Medicine

Background:

• Artificial intelligence (AI) offers potential efficiency gains in radiology and nuclear medicine, addressing workforce shortages.
• Slow adoption of AI-based clinical decision support systems (AI-CDSS) is attributed to limited model transparency.
• Explainable AI (XAI) aims to enhance clinician trust and acceptance by providing model transparency.

Purpose of the Study:

• To evaluate the impact of different XAI explanation types on radiologists' willingness to adopt AI systems.
• To assess the usefulness of XAI explanations in clinical decision-making for nuclear medicine.

Main Methods:

• Ten nuclear medicine radiologists used a simulated AI-CDSS for lung cancer TNM staging on PET/CT scans.
• Three XAI approaches were tested: input feature attribution, concept explanations, and global transparency.
• Radiologists rated adoption likelihood and explanation usefulness; qualitative interviews were also conducted.

Main Results:

• All XAI methods significantly increased adoption willingness compared to a black-box AI model (p < 0.05).
• Explanations were consistently useful for confirming or challenging AI staging recommendations (p < 0.001).
• Key factors influencing preferences included clinical relevance, error detection, decision support value, and a balance between explanation depth and usability.

Conclusions:

• XAI integration into nuclear medicine CDSS enhances radiologist acceptance and provides clinically relevant oversight information.
• These findings support XAI's role in facilitating the integration of AI tools into diagnostic workflows, aligning with regulatory expectations like the EU AI Act.