• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Monday, November 16, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Health

Multi-Agent AI System Restructures Radiology Reports and Flags Clinical Errors Locally

by
November 16, 2026
in Health
Reading Time: 6 mins read
0
Multi-Agent AI System Restructures Radiology Reports and Flags Clinical Errors Locally

Multi-Agent AI System Restructures Radiology Reports and Flags Clinical Errors Locally

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Medical researchers have demonstrated that a locally deployed multi-agent artificial intelligence architecture can systematically restructure narrative radiology reports into standardized anatomical categories while auditing them for clinical inconsistencies and reporting discrepancies. Developed and evaluated at the H. Lee Moffitt Cancer Center and Research Institute, the framework was tested on a retrospective cohort of 638 computed tomography reports of the chest, abdomen, and pelvis dictated by 15 board-certified radiologists across 2023 and 2024. The investigative team sought to address the long-standing challenge of heterogeneous formatting and unstandardized prose in radiology reporting, which frequently complicates communication with referring clinicians and limits secondary data extraction for quality monitoring and clinical decision support. Operating entirely behind the medical center’s institutional firewall, the automated pipeline unified deterministic sentence parsing and localized large language model reasoning within a single computational workflow that preserved original clinical phrasing without rewriting or summarizing content.

Radiology documentation typically incorporates findings and diagnostic impressions, but individual practitioners exhibit wide stylistic variation, ranging from free-form narrative paragraphs to diverse organ-specific headings. While conventional natural language processing techniques and newer generative models have previously demonstrated utility in summarizing reports or extracting discrete data entities, such approaches routinely condense, rephrase, or discard portions of the narrative text. This rewriting creates substantial risks of omitting vital clinical nuance or hallucinating unsupported medical assertions. To circumvent these failure modes, the researchers engineered a modular structuring pipeline that treated clinical text preservation as a foundational constraint. The resulting four-agent framework divided duties between deterministic rule sets and tiered open-source language models, ensuring that computationally lightweight components handled predictable linguistic patterns while reserving parameter-heavy reasoning engines for genuinely ambiguous anatomical contexts and comprehensive document-level quality surveillance.

The structuring pipeline processed the Findings section of each report after isolating section boundaries and applying custom segmentation algorithms designed to handle punctuation irregularities, numbered lists, and fragmented clauses such as brief stability notes. Agent 1 operated as a deterministic classifier driven by a domain-specific dictionary of approximately 600 regular expression patterns developed alongside a board-certified radiologist who did not author any cohort reports. When sentences fell outside these predefined regex rules, they were escalated to Agent 2, which utilized a quantized eight-bit version of LLaMA-3.1-8B-Instruct to infer organ-level associations using surrounding contextual cues. Sentences requiring deeper contextual interpretation, including complex device-related descriptions that necessitated simultaneous assignment to multiple anatomical domains, were routed directly to Agent 3. This third agent utilized a quantized four-bit DeepSeek-R1-Distill-Llama-70B model running locally to classify difficult phrasing and adjudicate sentences that Agent 2 placed into generic residual categories.

Quality assurance was managed by Agent 4, an independent reasoning module also powered by the quantized four-bit DeepSeek-R1-Distill-Llama-70B model. Operating at the whole-report level rather than sentence by sentence, Agent 4 evaluated internal consistency across the entire document, reconciling statements between the Findings and Impression sections as well as identifying contradictions within isolated sections. The agent specifically evaluated cross-sectional laterality mismatches, polarity contradictions, negation reversals, anatomical segment discrepancies, numerical or measurement unit conflicts, and documented communication gaps concerning potentially critical clinical findings. In addition, the quality assurance agent inferred patient gender from documented clinical history rather than anatomical references, systematically flagging discordant gender-specific procedures or anatomical designations. To ensure reliable automated integration, all language model agents generated structured outputs conforming to predefined JSON schemas, with Agent 4 equipped with an automated retry loop of up to five attempts for rare parsing failures.

Across the complete dataset of 638 radiology examinations containing 22,270 individual findings sentences, the tiered structuring architecture classified the overwhelming majority of narrative content without escalating to computationally intensive reasoning models. Agent 1 successfully categorized 18,493 sentences, representing 83.04 percent of the entire corpus, while Agent 2 classified 2,729 sentences, or 12.25 percent. Agent 3 processed 1,041 sentences, accounting for 4.67 percent of the total volume, leaving only seven sentences unclassified across the entire study. The researchers noted that these seven unclassified instances represented isolated section headers caused by spacing and line-break delimiters rather than true diagnostic phrasing. Individual radiologist writing styles significantly influenced agent utilization: clinicians adhering closely to conventional phrasing saw more than 90 percent of their sentences resolved by regex patterns alone, whereas reports from other practitioners required language model classification for more than 26 percent of their sentences.

In the automated quality assurance analysis, Agent 4 flagged 90 reports, representing 14.1 percent of the retrospective cohort, for suspected discrepancies or reporting oversights. The distribution of flagged issues varied substantially across individual radiologists, with flag rates per author ranging from 4.8 percent to 27.5 percent. Most detected anomalies were categorized under minor or major severity rather than acute emergencies: 44 reports exhibited minor clarity or consistency issues, 40 reports contained major contradictions carrying potential clinical consequences, and six reports were flagged for critical patient safety risks. The dominant category of identified error was section mismatches between the Findings and Impression or within sections, observed in 80 reports, or 12.5 percent of the dataset. Undocumented communication of potentially critical findings occurred in nine reports, while gender-anatomy discrepancies were identified in four reports across the entire study cohort.

Computational throughput was assessed on local hardware comprising a single NVIDIA H100 graphics processing unit and 12 central processing unit cores operating through the Ollama runtime platform. Across all 638 cases, the pipeline achieved a mean total processing time of 55.6 seconds per report, alongside a median duration of 39.3 seconds and an interquartile range spanning 26.6 to 60.2 seconds. The deterministic regex engine completed execution in an average of 0.539 seconds per report, while Agent 2 required an average of 1.310 seconds. In contrast, the reasoning tasks assigned to the 70-billion-parameter models constituted the vast majority of latency: Agent 3 averaged 28.618 seconds per report, and Agent 4 averaged 25.131 seconds. Runtimes varied markedly across author subsets, with mean durations ranging from 32.5 seconds to 265.5 seconds, reflecting a right-skewed distribution driven by occasional complex cases requiring extensive contextual adjudication by the larger reasoning models.

To evaluate clinical performance, two independent board-certified radiologists who were uninvolved in model design or report authorship conducted a blinded evaluation of a 45-report subset. This sample was intentionally enriched for system-flagged reports, containing 41 flagged documents and four unflagged controls to ensure rigorous auditing of error-detection capabilities. In evaluating anatomical restructuring, both reviewers agreed that 31 reports, representing 69 percent, were structured completely correctly, while three reports, or 7 percent, were restructured incorrectly due to at least one misclassified sentence. The reviewers disagreed on the remaining 11 cases, representing 24 percent, largely over organizational preferences concerning borderline anatomical boundaries. Crucially, both radiologists agreed across all 45 evaluated cases that the multi-agent system omitted no clinically important information and introduced zero fabricated assertions into the restructured output.

Regarding clinical review preferences, both evaluating radiologists agreed that the restructured document was about the same as the original narrative in 23 cases, better in one case, and worse in three cases, while disagreeing between better and about the same in 18 cases. Notably, the single report universally deemed better under restructuring was the sole examination in the cohort whose original Findings section was drafted without organ-based structure. In the quality assurance assessment, the reviewers confirmed that Agent 4 correctly identified section mismatches in 28 reports and correctly confirmed their absence in 11 reports, yielding an overall correct mismatch assessment rate of 87 percent alongside three false-positive flags. For critical communication tracking, the system correctly flagged five instances of undocumented communication and verified proper documentation in 38 cases, achieving a 96 percent accuracy rate alongside one false positive.

Reviewer agreement regarding Agent 4’s diagnostic reasoning and overall execution demonstrated robust qualitative performance across the enriched evaluation subset. The underlying clinical logic produced by the quality assurance agent was judged to be fully correct in 29 cases and mostly correct in six cases, with evaluator disagreement limited to distinguishing between fully and mostly correct classifications. Overall quality assurance performance was rated as excellent in 22 reports, good in two reports, and fair in seven reports by concordant review, with both radiologists assigning an excellent or good rating in 84 percent of evaluated cases and no report receiving a poor score. Interobserver reliability reflected high concordance for objective metrics, including Cohen’s kappa values of 0.87 for mismatch identification, 0.92 for critical communication tracking, and 1.00 for gender-anatomy assessments, while stylistic preferences and severity gradings exhibited moderate interobserver agreement.

The study authors highlighted several methodological limitations inherent to the retrospective investigation. The clinical evaluation cohort was constrained to 45 reports reviewed by two radiologists without sentence-level granular annotations or formal dispute adjudication, reflecting practical constraints in radiologist availability. Furthermore, because the evaluation sample was intentionally enriched with flagged cases, its performance metrics do not constitute unbiased population-level estimates for the general reporting pool. The system was optimized specifically for body CT protocols, meaning application to distinct anatomical regions or alternative modalities would necessitate tailored regex dictionaries and revised anatomical ontologies. Nevertheless, by demonstrating that deterministic rules can triage four-fifths of sentence-level classifications and preserve original narrative records, the researchers established the feasibility of local multi-agent architectures for integrated radiology standardization and automated quality surveillance.

Subject of Research: Local multi-agent artificial intelligence for radiology report structuring and quality assurance

Article Title: Multi-agent AI system for radiology report structuring and quality assurance with independent radiologist evaluation

Article References: Hartsock, I., Lam, C., Otteni, C., Qayyum, A., Gatenby, R., Araujo, C., & Rasool, G. (2026). Multi-agent AI system for radiology report structuring and quality assurance with independent radiologist evaluation. Frontiers in Radiology, 6, 1961438. https://doi.org/10.3389/fradi.2026.1961438

Image Credits: AI Generated

DOI: 10.3389/fradi.2026.1961438

Keywords: Artificial Intelligence, Radiology Reports, Computed Tomography, Multi-Agent Systems, Quality Assurance, Clinical Decision Support, Natural Language Processing, Large Language Models, Medical Informatics, Structured Reporting, Diagnostic Imaging, DeepSeek-R1, LLaMA-3, Moffitt Cancer Center, Patient Safety, Clinical Documentation

News Source: Ophelia Keating. (November 16, 2026). Multi-Agent AI System Restructures Radiology Reports and Flags Clinical Errors Locally. Scienmag.

Tags: Artificial IntelligenceClinical decision supportclinical documentationcomputed tomographyDeepSeek-R1Diagnostic imagingLarge Language ModelsLLaMA-3Medical InformaticsMoffitt Cancer Centermulti-agent systemsNatural Language ProcessingPatient Safetyquality assuranceradiology reportsstructured reporting
Share12Tweet7Share2ShareShareShare1

Related Posts

Patient-Reported Outcomes Show Broad Equity Differences Among Older Welsh Adults

Patient-Reported Outcomes Show Broad Equity Differences Among Older Welsh Adults

November 16, 2026
TyHGB Index Linked to Cardiometabolic Multimorbidity Across Three National Studies

TyHGB Index Linked to Cardiometabolic Multimorbidity Across Three National Studies

November 16, 2026

New Parent-Report Inventory Evaluates Four Core Emotional Competencies in Children

November 16, 2026

Caregiver-Child Cortisol Synchrony Linked to Lower Emotion Regulation in Preschool Children

November 16, 2026

POPULAR NEWS

  • Strengthening Patient Safety Culture in Fragile African Health Systems

    Strengthening Patient Safety Culture in Fragile African Health Systems

    29 shares
    Share 12 Tweet 7
  • Platform AI Assistance Linked to Higher Online Medical Consultation Service Quality

    29 shares
    Share 12 Tweet 7
  • Somali Pastoral Herd Sales Depend on Births and Theft Alongside Herd Size

    29 shares
    Share 12 Tweet 7
  • Silicon Oxide Composite Anode Retains High Capacity Over Three Hundred Cycles

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Patient-Reported Outcomes Show Broad Equity Differences Among Older Welsh Adults

TyHGB Index Linked to Cardiometabolic Multimorbidity Across Three National Studies

New Parent-Report Inventory Evaluates Four Core Emotional Competencies in Children

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.