Generative AI for automatic radiology report drafting: challenges, opportunities, and implementation requirements

Abstract

Radiology reporting is a high-stakes act of visual diagnosis, clinical synthesis, communication, and medico-legal documentation. Recent multimodal foundation models have made automatic report drafting technically plausible by combining image encoders, clinical language models, retrieval, longitudinal context, and interactive correction. This paper reviews the scientific and operational challenges that must be resolved before generative AI can be responsibly deployed in radiology reporting workflows. We argue that report drafting differs from conventional detection or classification because errors may appear as omitted findings, unsupported assertions, misleading recommendations, temporally incorrect comparisons, or fluent but non-evidential narrative. Implementation therefore requires multi-layer evaluation: image-level diagnostic performance; sentence- and finding-level factual consistency; clinical utility; uncertainty calibration; equity; cross-institutional generalization; and human-AI team performance under realistic workload. Opportunities include preliminary drafts, structured extraction, impression generation, quality assurance, communication standardization, teaching, and longitudinal disease tracking. However, deployment must address automation bias, vigilance fatigue, traceability, privacy, data shift, model updating, regulatory change control, and responsibility allocation. We propose an implementation framework centered on radiologist-in-the-loop supervision, auditable evidence links between images and report statements, continuous monitoring, locally validated acceptance criteria, and staged clinical evaluation. Generative AI should be treated not as an autonomous replacement for radiologists, but as an accountable drafting and reasoning substrate whose value depends on workflow design, measurement, and governance.