• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Wednesday, November 18, 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

Automated Agentic AI Workflow Reliably Diagnoses and Reports Macular Holes

by
November 18, 2026
in Health
Reading Time: 5 mins read
0
Automated Agentic AI Workflow Reliably Diagnoses and Reports Macular Holes

Automated Agentic AI Workflow Reliably Diagnoses and Reports Macular Holes

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Researchers have successfully validated an end-to-end agentic AI workflow designed to automate the diagnosis of macular holes from optical coherence tomography (OCT) images and generate structured, patient-friendly clinical reports. The study demonstrated that by coordinating a specialized retinal foundation model for image interpretation with a large language model for text synthesis, the architecture could achieve high levels of diagnostic accuracy while standardizing the communication of clinical findings. The workflow utilizes an orchestration agent to manage the information flow between these distinct components, ensuring that diagnostic decisions remain solely within the domain of the fine-tuned vision model, while the language model is restricted to interpreting those results for human readers. This separation of tasks addresses significant challenges in clinical AI, where transparency and the reliability of diagnostic conclusions are paramount for patient safety and workflow integration.

Macular hole (MH) is a significant vitreoretinal condition that often requires precise identification and monitoring to guide surgical and management decisions. While deep learning models have shown high efficacy in identifying retinal pathologies on OCT scans, most current applications operate as isolated diagnostic tools, requiring significant human intervention to bridge the gap between image analysis and the creation of clinical reports. As virtual and hybrid ophthalmology clinics expand to meet growing global patient volumes, the burden of interpreting images and translating findings into understandable clinic letters for patients and referring physicians has increased. The proposed agentic workflow seeks to bridge this divide by delegating the diagnostic task to a specialized computer vision model and the communication task to a large language model, effectively creating a more seamless bridge within the clinical pathway.

The methodological framework of this study centered on a modular architecture coordinated by a workflow orchestration agent. For the image diagnosis stage, the researchers fine-tuned the optical coherence tomography (OCT) branch of RETFound, a self-supervised retinal foundation model previously trained on millions of unlabeled images. By fine-tuning this model specifically for the binary classification of macular OCT scans as either macular hole or normal macula, the team leveraged the foundation model’s pre-existing feature representations to enhance efficiency and generalizability. Once the image diagnosis is performed, the result is relayed to the workflow agent, which constructs a structured prompt based on the diagnosis and standardized clinical documentation protocols used at Moorfields Eye Hospital. This structured prompt is then passed to a large language model, specifically GPT-4 via the OpenAI API, which generates a concise, professional summary covering the diagnosis, management considerations, and relevant risks.

Dataset management was a critical aspect of the study, ensuring rigorous training and validation. The researchers utilized a combination of several publicly available datasets for training and internal testing, including CHU, OCTDL, and OCTID for positive cases, and Kermany, OCTID, and OCTDL for normal images. To prevent data leakage and ensure realistic model performance, the study employed patient-level stratified splitting, ensuring that all images from an individual patient were assigned to only one subset. The training set comprised 691 MH and 897 normal images, while the validation set included 150 MH and 216 normal images. The internal test set contained 130 MH and 50 normal images. For external validation, the OIMHS and Noor Eye Hospital datasets provided an additional 563 and 451 images, respectively, allowing for an independent assessment of the model’s reliability across different scanners and acquisition protocols.

The model training performance demonstrated rapid convergence and stability throughout the 30-epoch fine-tuning process. On the held-out internal test set, the model achieved an accuracy of 100% with a 95% confidence interval ranging from 97.9% to 100%. The sensitivity and specificity were also recorded at 100%, with the area under the receiver operating characteristic curve (ROC AUC) reaching 1.00, indicating perfect discrimination on this specific dataset. No misclassifications were observed during the evaluation of these 180 test images, confirming the robustness of the foundation model when trained on homogeneous data. The training process remained consistent, with high F1 scores and ROC AUC metrics maintained throughout, suggesting that the fine-tuning approach successfully adapted the pre-trained features to the task of identifying macular holes.

External validation on the larger, independent dataset of 1,014 images provided a more challenging test of the model’s generalizability. The model achieved an overall accuracy of 98.8%, with 95% confidence intervals of 97.9% to 99.3%. Sensitivity for the detection of macular holes was 97.9%, and specificity for normal macula reached 100.0%. The ROC AUC was 0.999, with a confidence interval between 0.998 and 1.000. These results indicate that the fine-tuned foundation model maintains high diagnostic performance even when faced with data from different acquisition sources and patient populations. The confusion matrix for this stage showed 12 misclassified macular hole images and zero misclassified normal images, a performance profile that remains highly competitive for automated screening applications in vitreoretinal clinics.

Interpretability analysis formed an essential component of the validation, employing gradient-weighted class activation mapping (Grad-CAM) to visualize the model’s focus. For images with identified macular holes, the activation heatmaps consistently highlighted the hole and adjacent retinal layers, aligning with the expected clinical site of the pathology. Normal macula images, conversely, lacked this focused activation, showing instead diffuse, variable patterns that are consistent with the absence of localized retinal disruption. To confirm that these visualizations were grounded in learned parameters rather than arbitrary architectural features, the researchers conducted parameter randomization tests. The low Pearson correlations between attribution maps from the trained model and those from a randomized counterpart provided statistical evidence that the model’s explanations were robustly based on learned pathology-related features.

The clinical quality of the generated reports was assessed through independent review by two experienced ophthalmologists using a standardized three-point scale. Out of 40 randomly selected reports from the external validation set, 92.5% were found to be accurate and sufficient by both reviewers. Accuracy was defined as the correct diagnosis and clinically appropriate information, while sufficiency required the inclusion of management and risk-related details. While one reviewer rated every report as accurate and sufficient, the second reviewer identified 2.5% as accurate but insufficient and 5.0% as inaccurate, highlighting the potential for error in automated text generation.

In the broader context of clinical workflow automation, this study highlights the efficacy of a modular, agentic approach. Unlike general-purpose vision-language models, which have recently shown limitations in specialized ophthalmic tasks, this workflow maintains rigorous separation between interpretation and communication. By using a retina-specific foundation model for classification and a language model solely for synthesizing predefined clinical context, the architecture avoids the inherent risks of multimodal hallucinations or unreliable image interpretation by language models. This separation also facilitates modular updates; for instance, the diagnostic model could be upgraded or replaced with newer vision architectures without requiring a change to the report generation pipeline, making the system inherently more maintainable and adaptable to future clinical requirements.

The study also acknowledged several limitations that must be addressed before widespread clinical adoption. The model currently performs classification only on 2D OCT B-scans, potentially missing small or eccentric macular holes that might be identified through volumetric scan analysis. Furthermore, the dataset lacked clinical metadata regarding concomitant retinal conditions, which could potentially confound results in more complex patient scenarios. As the current focus was binary classification, the model does not currently stage the disease or provide prognostic data beyond the presence of a hole. Future iterations of this agentic AI workflow are expected to integrate volumetric data, perform automated image quality assessment, and include broader diagnostic categories, potentially incorporating data from electronic health records to provide personalized patient insights in a real-world virtual clinic environment.

Subject of Research: Clinical validation of an end-to-end agentic AI workflow for automated macular hole diagnosis and patient reporting.

Article Title: An Agentic AI Workflow for Automated Macular Hole Diagnosis and Structured Report Generation

Article References: Troyas, C., Ferro Desideri, L., Shah, N., Schumacher, I., Zinkernagel, M., & Anguita, R. (2026). An Agentic AI Workflow for Automated Macular Hole Diagnosis and Structured Report Generation. Ophthalmology and Therapy. https://doi.org/10.1007/s40123-026-01523-w

Image Credits: AI Generated

DOI: 10.1007/s40123-026-01523-w

Keywords: Ophthalmology, Artificial Intelligence, Optical Coherence Tomography, Macular Hole, Workflow Automation, RETFound, Deep Learning, Medical Imaging, Diagnostic AI, Large Language Models, GPT-4, Tele-ophthalmology, Clinical Report Generation, Virtual Clinics, Computer Vision, Foundation Models, Healthcare Technology, Retinal Pathology, Automated Diagnostics, Validation Study

News Source: Ophelia Keating. (November 18, 2026). Automated Agentic AI Workflow Reliably Diagnoses and Reports Macular Holes. Scienmag.

Tags: Artificial IntelligenceAutomated DiagnosticsClinical Report GenerationComputer Visiondeep learningdiagnostic AIfoundation modelsGPT-4Healthcare TechnologyLarge Language Modelsmacular holeMedical ImagingOphthalmologyoptical coherence tomographyRETFoundRetinal Pathologytele-ophthalmologyvalidation studyvirtual clinicsWorkflow Automation
Share12Tweet7Share2ShareShareShare1

Related Posts

Elite Luge Athletes Experience Seasonal Shifts in Body Composition and Biomarkers

Elite Luge Athletes Experience Seasonal Shifts in Body Composition and Biomarkers

November 18, 2026
Olive Oil Supplementation Remodels Plasma Lipids and Metabolites in Aged Tauopathy Mice

Olive Oil Supplementation Remodels Plasma Lipids and Metabolites in Aged Tauopathy Mice

November 18, 2026

Systematic Review Finds Long-Term Data Scarce for Chronic Neuropathic Pain Treatments

November 18, 2026

Cannabis Disorder Diagnosis Linked to Higher Hospitalization Risk in Heart Failure

November 18, 2026

POPULAR NEWS

  • Elite Luge Athletes Experience Seasonal Shifts in Body Composition and Biomarkers

    Elite Luge Athletes Experience Seasonal Shifts in Body Composition and Biomarkers

    29 shares
    Share 12 Tweet 7
  • Olive Oil Supplementation Remodels Plasma Lipids and Metabolites in Aged Tauopathy Mice

    29 shares
    Share 12 Tweet 7
  • Mitophagy Inhibitor Ac220 Enhances Magnolol Antitumor Action in Ewing Sarcoma

    29 shares
    Share 12 Tweet 7
  • Systematic Review Finds Long-Term Data Scarce for Chronic Neuropathic Pain Treatments

    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

Elite Luge Athletes Experience Seasonal Shifts in Body Composition and Biomarkers

Olive Oil Supplementation Remodels Plasma Lipids and Metabolites in Aged Tauopathy Mice

Mitophagy Inhibitor Ac220 Enhances Magnolol Antitumor Action in Ewing Sarcoma

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.