From Dictation to Structured Report: A Multicenter Evaluation of a Large Language Model–Assisted Dictation and Report Structuring Workflow in Radiological Reporting

Abstract The conventional dictation–transcription workflow may delay report draft readiness and is susceptible to operational errors, including the inadvertent deletion of normal baseline statements from the template. This study compared a large language model (LLM)-assisted speech-to-text conversion and template placement workflow with conventional transcription. In a controlled retrospective workflow simulation, 450 contrast-enhanced chest computed tomography (CT) examinations from three centers (one radiologist–transcriptionist pair each) were re-dictated. Using the same Turkish audio recording and standard template, 900 paired report drafts were produced. Drafts were scored on a 5-point Likert scale by the dictating radiologist, who was not informed of the workflow of origin; no independent reference-standard comparison was performed. The overall Findings-section score, reflecting perceived linguistic and structural quality, was 3.19 ± 0.49 in the conventional and 4.39 ± 0.27 in the LLM arm (difference 1.20; 95% confidence interval 1.15–1.24; p < 0.001), with the largest difference in spelling/grammar. Inadvertent deletion of normal baseline statements decreased from 192/450 (42.7%) to 32/450 (7.1%) drafts (risk difference 35.6%; p < 0.001). Time to draft readiness, including queue waiting in the conventional arm, was a median of 34.5 versus 1.00 min ( p < 0.001). LLM-assisted drafts had higher perceived linguistic and structural quality, fewer deletions of normal baseline statements, and earlier availability for review. Because clinical fidelity was not independently verified and correction time was not measured, these findings do not establish improved report accuracy or end-to-end efficiency and warrant prospective evaluation with independent evaluators, with radiologist final approval maintained.