Scholarlink

This study proposes a deep learning-based approach to classify dried droplet images of OM-85 under three conditions: no adulteration (control) and two levels of glucose adulteration (10% and 50%). The images were pre-processed using edge enhancement and segmentation techniques to remove background noise and isolate the droplet structure. The resulting images were then classified using the VGG-16 convolutional neural network. Classification performance improved significantly when using preprocessed images, with precision during training increasing from 49% to 91%. The model demonstrated particularly high precision for the 50% adulteration class, while most misclassifications occurred between the control and 10% categories, probably due to subtle visual similarities. Qualitative analysis revealed that structural features, such as the coffee ring, contributed to the successful classification. Our results support the potential of combining optical microscopy, image analysis and artificial intelligence for pharmaceutical quality control, as well as the development of rapid, non-invasive tools to detect adulteration and strengthen patient safety protocols in clinical and regulatory settings.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning-Based Detection of Dried OM-85 Droplets Adulterated with Glucose

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Carlos A. Martínez-Miwa,
  • Rocío Sánchez-Albores,
  • Josías N. Molina-Courtois,
  • Yojana J. P. Carreón,
  • Jorge González-Gutiérrez,
  • Mario Castelán

摘要

This study proposes a deep learning-based approach to classify dried droplet images of OM-85 under three conditions: no adulteration (control) and two levels of glucose adulteration (10% and 50%). The images were pre-processed using edge enhancement and segmentation techniques to remove background noise and isolate the droplet structure. The resulting images were then classified using the VGG-16 convolutional neural network. Classification performance improved significantly when using preprocessed images, with precision during training increasing from 49% to 91%. The model demonstrated particularly high precision for the 50% adulteration class, while most misclassifications occurred between the control and 10% categories, probably due to subtle visual similarities. Qualitative analysis revealed that structural features, such as the coffee ring, contributed to the successful classification. Our results support the potential of combining optical microscopy, image analysis and artificial intelligence for pharmaceutical quality control, as well as the development of rapid, non-invasive tools to detect adulteration and strengthen patient safety protocols in clinical and regulatory settings.