Ophthalmology and therapy · 2026
An Agentic AI Workflow for Automated Macular Hole Diagnosis and Structured Report Generation.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Troyas Carla, Ferro Desideri Lorenzo, Shah Neil, Schumacher Inès, Zinkernagel Martin, Anguita Rodrigo
Аннотация
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 foundation model, was fine-tuned for binary classification of macular OCT scans as MH or normal macula. Training, validation, and internal testing were performed using multiple public OCT datasets with patient-level stratified splitting, followed by external validation on two independent datasets. The performance of the fine-tuned foundation model was evaluated using accuracy, sensitivity, specificity, and receiver operating characteristic area under the curve (ROC AUC). For each new OCT image, the AI agent used the trained RETFound model to classify the image, received the diagnostic output, incorporated it into a structured prompt, and called the large language model (GPT-4) to generate the clinical report.
RESULTS: On the internal test set (n = 180), the model achieved an accuracy of 100% (95% confidence interval (CI), 97.9-100%) with an AUC of 1.00. On external validation (n = 1014), accuracy was 98.8% (95% CI 97.9-99.3%), sensitivity was 97.9% (95% CI 96.3-98.8%), specificity was 100% (95% CI 99.2-100.0%), and AUC was 0.999 (95% CI 0.998-1.000). Gradient-weighted Class Activation Mapping (Grad-CAM) analysis showed that predictions were driven by relevant retinal regions, particularly the macular hole and adjacent layers. The end-to-end agentic AI workflow generated concise and standardized patient-friendly reports for all samples, without the language model influencing diagnostic decisions.
CONCLUSIONS: This study demonstrates the feasibility of an end-to-end AI workflow coordinated by a workflow orchestration agent for robust OCT-based MH diagnosis with a large language model for structured report generation. By separating image-based diagnosis from language generation, this approach offers a practical and scalable solution for virtual vitreoretinal clinics.
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 foundation model, was fine-tuned for binary classification of macular OCT scans as MH or normal macula. Training, validation, and internal testing were performed using multiple public OCT datasets with patient-level stratified splitting, followed by external validation on two independent datasets. The performance of the fine-tuned foundation model was evaluated using accuracy, sensitivity, specificity, and receiver operating characteristic area under the curve (ROC AUC). For each new OCT image, the AI agent used the trained RETFound model to classify the image, received the diagnostic output, incorporated it into a structured prompt, and called the large language model (GPT-4) to generate the clinical report.
RESULTS: On the internal test set (n = 180), the model achieved an accuracy of 100% (95% confidence interval (CI), 97.9-100%) with an AUC of 1.00. On external validation (n = 1014), accuracy was 98.8% (95% CI 97.9-99.3%), sensitivity was 97.9% (95% CI 96.3-98.8%), specificity was 100% (95% CI 99.2-100.0%), and AUC was 0.999 (95% CI 0.998-1.000). Gradient-weighted Class Activation Mapping (Grad-CAM) analysis showed that predictions were driven by relevant retinal regions, particularly the macular hole and adjacent layers. The end-to-end agentic AI workflow generated concise and standardized patient-friendly reports for all samples, without the language model influencing diagnostic decisions.
CONCLUSIONS: This study demonstrates the feasibility of an end-to-end AI workflow coordinated by a workflow orchestration agent for robust OCT-based MH diagnosis with a large language model for structured report generation. By separating image-based diagnosis from language generation, this approach offers a practical and scalable solution for virtual vitreoretinal clinics.