What happened
A research team led by the University of Minnesota has demonstrated that a hybrid open-source optical character recognition (OCR) can accurately extract genomic recurrence scores from scanned Oncotype DX breast cancer reports. The study, published in Cancer Causes & Control, utilized a combination of Tesseract and EasyOCR engines to process 675 reports, achieving a 97% agreement rate with manual human abstraction. The automated system significantly outperformed the speed of traditional manual registry entry, which often faces 12 to 18-month delays.
The research team, including Qianyun Luo and Schelomo Marmor, developed a 'hybrid OCR' (H-OCR) to address the in cancer registries. Genomic test results, such as the Oncotype DX recurrence score, are frequently stored in electronic health records (EHRs) as unstructured scanned images rather than machine-readable data, requiring manual transcription.
The H-OCR uses EasyOCR for its deep-learning-based text detection and Tesseract as a fallback engine for character recognition. This hybrid approach was designed to leverage EasyOCR's ability to localize circled annotations on reports while utilizing Tesseract's speed and character recognition capabilities.
In a validation set of 675 reports, the H-OCR achieved 97% agreement with manual reference standards, outperforming the 91% accuracy rate of conventional registry abstraction. The automated process completed the task in approximately 4.9 hours, compared to the months-long delays typical of current manual workflows.
The study found that the 's confidence scores effectively identified potential errors, suggesting that future implementations could use these scores to flag uncertain extractions for human review, thereby minimizing the risk of clinically significant misclassification.
Source details: bioengineer.org ↗
Why it matters
The current reliance on manual transcription for genomic data creates a significant bottleneck in cancer research and precision oncology. By automating the extraction of structured data from unstructured PDFs, this open-source approach allows health systems to populate cancer registries in near real-time. This shift is critical for improving the quality of real-world evidence, enabling faster comparative effectiveness research, and providing the structured data necessary to train future AI models in oncology. Because the tools are open-source and can operate within HIPAA-compliant environments, this method offers a low-cost, reproducible solution for resource-constrained institutions to modernize their data infrastructure without proprietary software.
Genomic biomarkers are essential for precision oncology and quality measurement, yet their utility is currently limited by the time it takes to make them available in research databases. Reducing this allows for more timely evidence generation.
The use of open-source tools like Tesseract and EasyOCR ensures that the solution is accessible to smaller or resource-constrained cancer centers that may lack the budget for expensive, proprietary medical data extraction software.
The study demonstrated that the benefits of this automated capture are broadly applicable across diverse patient populations, as the researchers found that patient demographics and clinical factors did not significantly predict extraction errors.
By converting unstructured binary objects into structured data, this provides the high-quality, longitudinal data necessary for training and refining future AI and machine learning models in the oncology space.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
What is the best response when AI Models Explained makes a mistake in production?
What to watch next
Future research will need to address the scalability of this beyond the standardized Oncotype DX format. The authors noted that somatic next-generation sequencing reports, which exhibit greater heterogeneity across vendors, present a more complex challenge for automated extraction. Additionally, while the study validated the pipeline against a retrospective dataset, prospective testing within live clinical registry workflows is required to confirm its reliability in real-world, high-volume settings.
The researchers explicitly identified the need to test the on more heterogeneous document types, such as somatic next-generation sequencing reports, which vary significantly by vendor and format.
The study was conducted at a single health system; future work must evaluate the 's performance across multiple institutions to ensure it remains robust against variations in scan quality, fax resolution, and different EHR configurations.
Prospective integration into live registry workflows is the next logical step to determine if the system maintains its accuracy and efficiency when processing incoming clinical data in real-time.