Truace
Subscribe

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

Introduction: Current AI models for optical coherence tomography (OCT) primarily focus on diagnosis as a standalone task. We developed and validated an end-to-end agentic AI workflow that integrates automated macular hole (MH) diagnosis from OCT images with structured patient-friendly report generation. Methods: The proposed sequential workflow coordinated by a workflow orchestration agent consisted of an image diagnosis stage and a language generation step. The OCT version of RETFound, a self-supervised retinal…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2232· cite & verify
An Agentic AI Workflow for Automated Macular Hole Diagnosis and Structured Report Generation
Retinal Pigment Epithelium (RPE) Cells by NIH Image Gallery. Public domain

In brief

Researchers built a two-stage workflow coordinated by an orchestration agent. A fine-tuned version of the RETFound retinal foundation model first classifies each OCT scan as macular hole or normal macula. The result is then placed into a structured prompt, and GPT-4 writes a clinical report.

The design matters because it keeps diagnosis separate from text generation, so the language model does not decide the diagnosis. The authors describe it as a practical, scalable option for virtual vitreoretinal clinics, but the evidence comes from public datasets and a binary task, so real-world clinical performance is not yet shown.

Main points

  1. The workflow combines an image diagnosis stage with a language generation step, coordinated by an orchestration agent.
  2. On the internal test set of 180 scans, accuracy was 100% with an AUC of 1.00.
  3. On external validation of 1014 scans, sensitivity was 97.9%, specificity was 100%, and AUC was 0.999.
  4. Grad-CAM analysis showed predictions were driven by relevant retinal regions, particularly the macular hole and adjacent layers.

The gain

A fine-tuned retinal foundation model detected macular holes on external OCT datasets with high accuracy, while a language model produced standardized patient-friendly reports without influencing the diagnosis.

The rundown

The diagnostic stage used the OCT version of RETFound, a self-supervised retinal foundation model, fine-tuned for binary classification of macular OCT scans as macular hole or normal macula. Training, validation and internal testing used multiple public OCT datasets with patient-level stratified splitting.

Performance was checked on two independent external datasets using accuracy, sensitivity, specificity and ROC AUC. Grad-CAM heat maps were used to examine which image regions drove predictions.

For each new image, the agent passed the classifier's output into a structured prompt and called GPT-4 to write the report. The study reports that reports were concise and standardized for all samples.

Sources

  1. Peer-reviewedOphthalmology and Therapy2026-10-23

The debate