OCR-Based Pipeline for Table Extraction from Spine Dual X-Ray Absorptiometry Reports

Identifier:  TFM:2535

Handle: https://hdl.handle.net/20.500.11797/TFM2535

Authors:  Leal Rodríguez, Samuel

Abstract:
This thesis presents an OCR-based pipeline for automatically extracting structured data from Dual X-Ray Absorptiometry (DXA) reports used in osteoporosis diagnosis. The system uses fixed coordinate cropping and color-based segmentation to identify three key tables containing patient demographics, examination metadata, and bone mineral density measurements. Using Surya OCR for text recognition and spatial coordinate analysis for table reconstruction, the pipeline achieved excellent table detection accuracy on 100 standardized Hologic DXA images from 2020-2025. The approach prioritizes computational efficiency over generalizability, converting unstructured medical images into machine-readable CSV format for research applications and electronic health record integration.