Multi-modal volumetric concept activation to explain detection and classification of metastatic prostate cancer on PSMA-PET/CT

AI explanations can be made understandable for medical imaging: The study successfully demonstrated that concept activation—a technique that translates AI decision-making into human-interpretable concepts—can explain how neural networks detect and classify cancer in medical scans, making AI behavior more transparent to clinicians.

Different data types serve different purposes in AI diagnosis: The research found that when analyzing PET/CT scans, the AI system used CT images primarily to identify anatomical location while relying on PET images to build confidence in cancer detection, showing that explainable AI can reveal how and why algorithms combine multiple information sources.

Explainability can help reduce diagnostic errors: Local explanations generated by this method showed promise in distinguishing between correct cancer detections and false alarms, suggesting that understanding AI reasoning could help clinicians catch and correct mistakes before they affect patient care.

This approach works for complex, real-world medical data: The feasibility demonstration used actual metastatic prostate cancer patients and three-dimensional volumetric imaging data, indicating that explainability techniques can scale beyond simple datasets to clinically relevant scenarios.

Responsible AI in healthcare requires both performance and interpretability: Achieving 80% sensitivity with manageable false positive rates while simultaneously providing explanations demonstrates that medical AI systems can be both accurate and interpretable—supporting safer clinical adoption.