Improving tabular data extraction in scanned laboratory reports using deep learning models.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Clinical Laboratory Science

Background:

• Medical laboratory testing is vital for healthcare decisions.
• Current methods often use faxed reports, delaying data access.
• Efficient extraction of lab data from scanned documents is needed.

Purpose of the Study:

• To develop a deep learning-based Optical Character Recognition (OCR) method for identifying and extracting lab testing results from scanned reports.
• To improve the timeliness and accuracy of clinical data availability.

Main Methods:

• A two-stage approach was used: table detection and table recognition.
• Deep learning models DETR R18 and YOLOv8s were evaluated for table detection.
• PaddleOCR and the encoder-dual-decoder (EDD) model were compared for table recognition.
• 650 tables from 632 reports were annotated for training and evaluation.

Main Results:

• Fine-tuned DETR R18 achieved superior performance in table detection (AP50: 0.774).
• Fine-tuned EDD demonstrated strong performance in table recognition (TEDS score: 0.815).
• The integrated OCR pipeline achieved a TEDS score of 0.699 and a TEDS structure score of 0.764.

Conclusions:

• A novel OCR pipeline using state-of-the-art deep learning models was developed for scanned clinical documents.
• The pipeline effectively extracts tabular lab testing data, enhancing clinical data analysis.
• This approach has significant implications for improving clinical decision-making through accessible data.