PS47: From Pixels to Clinical Implementation: AI-Enabled Digital Tools in Precision Oncology

Artificial intelligence (AI) and machine learning–based image analysis are transforming precision oncology by enabling the automated derivation of biomarker and clinical endpoints from histopathology and radiologic imaging. Two prominent applications-digital pathology and AI-RECIST-apply AI-driven computational tools to quantify biomarker expression and assess tumor response, respectively. These approaches have the potential to reduce inter- and intra-evaluator variability, improve diagnostic precision, and streamline clinical workflows. Recent milestones, such as the FDA Breakthrough Device Designation granted to the Quantitative Continuous Scoring (QCS) platform-the first AI-driven companion diagnostic-highlight growing regulatory recognition of these technologies. Parallel efforts, including initiatives led by Friends of Cancer Research, further reflect increasing momentum toward the adoption of AI-derived digital endpoints in oncology research. Despite these advances, the development of AI-driven digital tools faces substantial statistical, clinical, and operational challenges. A central issue is limited data availability: many tools are trained and initially validated using early-phase clinical trial or preclinical datasets, where sample sizes are small, patient populations are heterogeneous and may differ meaningfully from those in late-phase registrational studies. These constraints complicate performance evaluation, uncertainty quantification, and assessment of generalizability. For AI-RECIST in particular, additional challenges include heterogeneity in imaging acquisition protocols, lesion selection and longitudinal tracking variability, handling non-measurable or missing lesions, and alignment with RECIST criteria originally designed for human readers. Addressing these challenges requires carefully designed validation strategies, including cross-validation, independent dataset evaluation, and benchmarking against standard-of-care assessments, alongside formal evaluation of bias and robustness. Model interpretability and regulatory evaluation present further complexity. Even when AI tools demonstrate strong empirical performance, they are often perceived as "black boxes," raising questions about transparency, performance monitoring, and integration into clinical decision-making. Regulatory evaluation emphasizes analytical validity, clinical relevance, and reproducibility, yet expectations for AI-derived endpoints continue to evolve. This session brings together industry experts-including statisticians, pathologists, and radiologists-to share hands-on experience in statistical methodology and clinical development for AI-based digital pathology and RECIST tools. The focus is on integrating tool development into the clinical development timeline to enable reliable use prior to registrational studies. Seasoned FDA reviewer(s) will provide regulatory perspectives. Through practical examples and discussion, the session highlights both the opportunities and challenges of validating and deploying AI-derived imaging endpoints, fostering rigorous, responsible, and cross-disciplinary adoption of novel digital tools in precision oncology.