Evidence-Linked Radiology Reporting: A Human-Supervised Reference Architecture for Structured Imaging Intelligence

A proposed reference architecture aims to extract structured data from free-text radiology reports, linking findings to image evidence while keeping radiologists in supervisory control, according to a paper posted on arXiv [1]. The framework, described as a human-supervised, evidence-linked system, targets a longstanding problem in medical imaging: measurements, lesion identities, prior comparisons, and terminology often remain locked in narrative prose or scattered across picture archiving and communication systems, radiology information systems, and electronic health records [1][2]. The authors position the system not as an autonomous report generator but as a structured intelligence layer that supports reviewed reporting, longitudinal comparison, and clinical data reuse [1][2]. The architecture combines exam-specific templates, speech-to-structure processing, measurement and segmentation capture, and controlled AI-assisted drafting [1][2]. Interoperability relies on established standards including DICOM, DICOM Structured Reporting, DICOM Segmentation, HL7 FHIR, RadLex, SNOMED CT, LOINC, and UCUM [1][2]. The paper also addresses modality-specific deployment considerations, clinical safety risks, validation requirements, cybersecurity, privacy, quality management, and regulatory boundaries for AI-assisted radiology reporting systems [1][2]. Health informatics, the broader discipline in which such architectures sit, applies computational techniques to improve the communication and management of medical information, drawing on fields from software engineering to data science [4]. The proposed system’s AI-assisted drafting component intersects with advances in large language models, which are neural networks trained on vast text corpora for generation, summarization, and parsing tasks [3]. Benchmark evaluations for these models attempt to measure reasoning, factual accuracy, alignment, and safety — concerns that parallel the validation and clinical safety risks flagged in the radiology reporting proposal [3][1]. The paper emphasizes that the architecture is designed for enterprise imaging integration with PACS, RIS, EHR, analytics, and registry workflows, reinforcing its role as a governance-aware layer rather than a standalone application [1][2].