Publication Date
Spring 2026
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
First Advisor
Mark Stamp
Second Advisor
Thomas Austin
Third Advisor
Navrati Saxena
Keywords
pharmaceutical labeling, FDA compliance, generative AI, structured product labeling (SPL), image quality assessment
Abstract
This work describes a framework for generating and evaluating FDA-compliant pharmaceutical labels using an SPL XML parser to extract drug metadata and label text through LOINC mappings, followed by rendering through Gemini models from Google. When a sample label is provided, Gemini 2.5 Pro extracts visual style features for brand consistency; otherwise, an FDA Q&A session pre-fills the template fields from the XML document. Using Gemini 3.1 Flash, label images are generated through a constraint prompt with required information including the NDC barcode. The FDA compliance validation module evaluates labels against 21 CFR Part 201 requirements by applying 12 pre-render field presence tests, and 3 post-render verifications (OCR text verification barcode scanning and minimum font size), resulting in a weighted compliance score. For assessing visual quality of the generated labels, we propose a multi-metric visual evaluation model based on LPIPS perceptual similarity, SSIM, MS- SSIM, color histogram intersection, layout similarity through edges, and semantic text embedding similarity. Experiments on actual DailyMed SPL labels show promising results for producing visually accurate pharmaceutical labels using a robust evaluation process.
Recommended Citation
Deore, Shreyas Satish, "Quantitative Evaluation and Generation of FDA-Compliant Pharmaceutical Labels" (2026). Master's Projects. 1776.
DOI: https://doi.org/10.31979/etd.agw4-n852
https://scholarworks.sjsu.edu/etd_projects/1776
