Ocular Biometry OCR: a machine learning algorithm leveraging optical character recognition to extract intra ocular...

Area of Science:

• Ophthalmology
• Medical Informatics
• Computer Vision

Background:

• Ocular biometry measurements are crucial for predicting ophthalmic diseases.
• Manual data extraction from PDF reports is time-consuming and labor-intensive.
• Lack of structured data hinders the use of biometry in machine learning models.

Purpose of the Study:

• To develop and evaluate an automated system, Ocular Biometry OCR, for extracting ocular biometry measurements from optical biometry reports.
• To compare the performance of PaddleOCR and Gemini models in extracting this data.
• To facilitate the use of biometry data in machine learning for eye disease prediction.

Main Methods:

• Utilized two OCR models, PaddleOCR and Gemini, to process 2,965 Lenstar, 104 IOL Master 500, and 3,616 IOL Master 700 reports.
• Developed a text extraction pipeline involving cropping, OCR, key-value pair processing, error correction, and result selection.
• Evaluated model performance using detection, similarity, and agreement scores, with IOLM 500 data annotated for ground truth.

Main Results:

• The Ocular Biometry OCR system demonstrated high performance across all tested devices.
• Detection scores ranged from 0.990 to 1.000, similarity scores from 0.995 to 0.999, and agreement scores from 0.985 to 0.999.
• Both models showed excellent agreement, with PaddleOCR exhibiting slightly superior performance in qualitative and quantitative analysis.

Conclusions:

• Automated extraction of ocular biometry data using OCR is feasible and highly accurate.
• The Ocular Biometry OCR system significantly reduces manual data entry efforts.
• This technology can accelerate the development of machine learning-based ophthalmic disease prediction models.