Abstract
Optical Coherence Tomography (OCT) is the gold standard for diagnosing retinal diseases, yet clinical data often remains trapped in static, proprietary image formats. This study introduces a deep learning-based pipeline to transform these non-standardized reports into structured, interoperable data. By integrating a custom OCR engine with a spatial "Area Explorer," the system achieves 99.67% accuracy across nearly half a million clinical reports. The resulting data is converted into the OMOP Common Data Model (CDM), facilitating large-scale longitudinal research and cross-institutional interoperability.
Introduction
In modern ophthalmology, the "data silo" problem significantly hinders clinical research and patient care. While OCT devices provide critical diagnostic information, they typically output reports as static PDFs or JPEGs. Because manufacturers use proprietary formats, interoperability between hospital systems is virtually non-existent. Clinicians must manually review historical PDFs to track disease progression, and researchers face immense manual data entry burdens. This paper proposes an automated Extract, Transform, and Load (ETL) strategy to unlock this trapped information.
Methodology: The Deep Learning Pipeline
The researchers developed a modular software suite to automate the extraction and standardization process. The architecture is designed to handle the spatial complexity of medical reports while ensuring high-precision data recovery.
The pipeline consists of three primary components:
- Area Explorer: This module utilizes configuration files to identify specific regions of interest (ROIs) on a report. It distinguishes between numeric values and tomographic images, providing the necessary spatial context.
- Value Reader (OCR): To overcome the limitations of standard OCR tools on specialized medical fonts, a custom engine was built. It employs a ResNet backbone for feature extraction and a Bidirectional LSTM for sequence modeling, specifically optimized for numeric measurements.
- Finding Classifier: This deep learning model analyzes the tomographic images within the report to identify pathological features, such as subretinal fluid or macular edema.
Data Extraction and Validation
The system was trained and validated using a combination of synthetic and real-world clinical data:
- Synthetic Training: To avoid labeling fatigue, 1 million synthetic images were generated to simulate various font sizes, contrasts, and noise levels typical of medical scans.
- Clinical Evaluation: A set of 300 OCT reports was manually labeled by experts to serve as a ground truth.
- Large-Scale Deployment: The pipeline processed an archive of 433,175 OCT reports from Seoul National University Bundang Hospital, spanning the years 2006 to 2019.
Key Results and Insights
The Value Reader achieved a remarkable 99.67% accuracy in converting image-based numbers to structured text. This high precision allowed for the successful processing of over 13 years of historical data, creating a longitudinal dataset that enables clinicians to filter and sort patient progress effortlessly.
A critical insight from this work is the importance of structural context. By knowing exactly where a number is located (via the Area Explorer), the system understands its clinical meaning—for example, distinguishing between central macular thickness and retinal nerve fiber layer thickness. Furthermore, by targeting the OMOP Common Data Model (CDM), the authors enable ophthalmic data to be analyzed alongside broader clinical datasets, such as pharmacy and genomic records, in multi-center studies.
Conclusion and Future Directions
While the current system demonstrates high throughput and precision, it is primarily designed for retrospective batch processing. Future work aims to develop real-time integration with Electronic Health Records (EHR) and expand support to a wider variety of device manufacturers and exam types, such as corneal topography. Despite these challenges, this strategy represents a significant step toward breaking down data silos in ophthalmology and enabling truly interoperable medical research.
Summary updated