Transfer learning and data augmentation for glucose concentration prediction from colorimetric biosensor images.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Analytical Chemistry

Background:

• Accurate glucose concentration monitoring is crucial for diabetes management.
• Traditional methods for glucose detection can be invasive or complex.
• Image-based sensing offers a promising non-invasive alternative.

Purpose of the Study:

• To develop a deep learning algorithm for predicting glucose concentrations using colorimetric paper sensor (CPS) images.
• To evaluate the efficacy of transfer learning and data augmentation in improving prediction accuracy.
• To simplify the glucose prediction process by eliminating the need for manual feature extraction.

Main Methods:

• Utilized an image dataset of CPS treated with varying glucose concentrations.
• Applied transfer learning by adapting four deep learning models (ResNet50, ResNet101, GoogLeNet, VGG-19).
• Implemented data augmentation techniques to address the need for large training datasets.

Main Results:

• GoogLeNet achieved the highest prediction accuracy (R² = 0.994).
• Deep learning models significantly outperformed traditional machine learning approaches (p < 0.001).
• Data augmentation with 20% of the training data yielded comparable prediction error to using the full dataset.

Conclusions:

• A novel deep learning approach enables accurate glucose prediction from CPS images.
• Transfer learning and data augmentation effectively enhance prediction performance and reduce data requirements.
• This method offers a simplified, image-based approach for glucose concentration determination.