Explainable AI in Nuclear Medicine: Clinical Decision Support from Implementation to Adoption

Abstract:
Artificial intelligence (AI) offers significant potential to alleviate workforce pressures and facilitate personalised clinical decision-making; however, its clinical adoption remains limited. A lack of model transparency hinders user trust and compromises the human oversight mandated by legislative frameworks such as the EU AI Act. This thesis investigates the implementation and adoption of eXplainable AI (XAI) across the nuclear medicine clinical pathway, employing a people-centred approach that utilises clinical perspectives to guide technical development.

First, a task-based user study evaluated radiologists' preferences across a proposed Hierarchy of AI Interpretability during PET-CT lung cancer staging. The findings demonstrated that XAI significantly increased adoption willingness compared to black-box systems, revealing a fundamental trade-off between explanation depth and clinical usability. Second, a multimodal deep learning framework was developed to predict treatment outcomes for metastatic castration-resistant prostate cancer patients undergoing Radium-223 therapy. Incorporating a domain-adapted vision encoder alongside text reports and tabular variables, the pre-treatment neural networks successfully outperformed conventional statistical baselines. Finally, correlational and causal explainability pipelines were applied to the Radium-223 outcome prediction model and critically compared. Concept Erasure successfully isolated clinical phenotypes within the latent space, providing faithful causal explanations where input-space perturbations struggled with cross-correlated multimodal data.

Despite the enormous diagnostic and prognostic potential offered by AI, this thesis demonstrates that high-performing AI alone may be insufficient for clinical acceptance; clinicians prefer transparent explanations to confidently adopt decision support systems. The initial user study established which type of XAI is suitable for medical professionals, whereas the final investigation determined what type of XAI is suitable for medical data. By uniting these clinical user requirements with causal technical implementations, this work establishes a foundation for trustworthy AI tools that enable mandatory human oversight and advance personalised patient selection in nuclear medicine.