Biological and Radiological Dictionary of Radiomics Features: Addressing Understandable AI Issues in Personalized Prostate Cancer; Dictionary Version PM1.0

AI interpretability matters in healthcare: The paper demonstrates that AI models predicting prostate cancer risk can be made understandable to doctors by creating a "dictionary" linking AI features to actual biological and radiological concepts. The best-performing model achieved 78% accuracy while remaining interpretable to medical professionals, showing that accuracy and explainability don't have to be mutually exclusive.

Standardized frameworks bridge the AI-medicine gap: By establishing shared terminology between AI specialists and radiologists (the PM1.0 dictionary), the research addresses a key responsible AI challenge—ensuring that non-technical healthcare professionals can understand, validate, and trust AI recommendations rather than treating them as "black boxes."

Combining multiple data sources improves reliability: Using three different MRI imaging sequences (T2WI, DWI, and ADC) together significantly outperformed single-sequence methods (p-value<0.05), suggesting that responsible AI deployment in personalized medicine should integrate multiple data sources to reduce errors and improve clinical confidence.

Specific, measurable features enable accountability: The identified key predictive features (such as lesion intensity, heterogeneity, shape metrics, and texture patterns) are concrete and traceable, allowing clinicians to audit why the AI made specific predictions and verify whether those reasons align with established medical knowledge about prostate cancer.