Abstract—Digitizing pharmaceutical labels using Optical Character Recognition (OCR) can improve medication safety and hospital workflow efficiency, but cross-hospital deployment remains challenging due to variations in label templates, printing formats, and limited labeled medical data. This study evaluates text detection and recognition models for Thai pharmaceutical label OCR under cross-hospital conditions using a three-tier evaluation protocol: source-hospital data, mixed-hospital data, and unseen hospitals. The results show clear architectural differences in cross-hospital robustness. The PaddleOCR recognition model achieves the best cross-hospital performance with 80.47% accuracy on unseen hospitals, while Transformer-based TrOCR variants degrade substantially under domain shift, achieving only 30–52% accuracy. Text detection shows relatively stable transfer across hospitals with a moderate drop of −10.68%, whereas recognition performance varies notably across architectures and training strategies. Conservative fine-tuning improves recognition accuracy by 14.85% over the base model while preserving cross-hospital generalization. These findings suggest that pre-training diversity and architectural inductive biases are more important than model scale for OCR deployment in limited-data medical settings.
 
Keywords—cross-hospital generalization, medical document processing, optical character recognition, pre-training quality, Thai script
 
Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).