Free papers detail | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Development of Foundation Model–Driven Full-Process 3D OCT Diagnostic System and Clinical Utility Evaluation

Authors
Jinze Zhang, Jian Zhong, Jiaxiong Li, Xiaoying Tang, Peng Xiao

Presenting
Jinze Zhang

PURPOSE:
Optical Coherence Tomography (OCT)-based retinal diagnosis involves multi-step workflows, including quality control, anomaly detection, and multi-slice interpretation. Existing single-task, single-slice AI models are insufficient for such complexity. This study aimed to develop a foundation model-driven 3D OCT system for full-process diagnosis, validate it on real-world data, and assess its clinical utility.

PURPOSE:

Optical Coherence Tomography (OCT)-based retinal diagnosis involves multi-step workflows, including quality control, anomaly detection, and multi-slice interpretation. Existing single-task, single-slice AI models are insufficient for such complexity. This study aimed to develop a foundation model-driven 3D OCT system for full-process diagnosis, validate it on real-world data, and assess its clinical utility.

METHODS:
We developed the Full-process OCT-based Clinical Utility System (FOCUS), integrating image quality assessment (EfficientNetV2-S), anomaly detection, and multi-disease classification (VisionFM), followed by a unified adaptive aggregation strategy for patient-level 3D diagnosis. The system was trained and tested on 3,300 patients (40,672 slices) and externally validated on 1,336 patients (19,798 slices) from multiple centers. Human–AI comparison experiments were conducted to evaluate diagnostic consistency across clinical roles. Decision Curve Analysis (DCA) was used to assess clinical net benefit.

METHODS:

We developed the Full-process OCT-based Clinical Utility System (FOCUS), integrating image quality assessment (EfficientNetV2-S), anomaly detection, and multi-disease classification (VisionFM), followed by a unified adaptive aggregation strategy for patient-level 3D diagnosis. The system was trained and tested on 3,300 patients (40,672 slices) and externally validated on 1,336 patients (19,798 slices) from multiple centers. Human–AI comparison experiments were conducted to evaluate diagnostic consistency across clinical roles. Decision Curve Analysis (DCA) was used to assess clinical net benefit.

RESULTS:
FOCUS achieved high performance in quality assessment (F1=99.01%), anomaly detection (F1=97.46%), and classification (F1=94.39%), outperforming comparison methods. External validation showed consistent performance (F1: 90.22%–95.24%) across centers. FOCUS surpassed ophthalmic technicians in anomaly detection and reached retinal specialist-level performance in diagnosis, with improved efficiency. DCA demonstrated positive net benefit across a wide range of risk thresholds.

RESULTS:

FOCUS achieved high performance in quality assessment (F1=99.01%), anomaly detection (F1=97.46%), and classification (F1=94.39%), outperforming comparison methods. External validation showed consistent performance (F1: 90.22%–95.24%) across centers. FOCUS surpassed ophthalmic technicians in anomaly detection and reached retinal specialist-level performance in diagnosis, with improved efficiency. DCA demonstrated positive net benefit across a wide range of risk thresholds.

CONCLUSIONS:
The FOCUS system automates the analytical workflow from image quality control to 3D OCT diagnosis. It demonstrates stable diagnostic capability and good generalization performance in identifying various retinal diseases, consistently yielding a positive clinical net benefit. This provides a novel technical pathway for Foundation model-based intelligent diagnosis of fundus diseases using OCT.