MP-Net

MP-Net quantifies fluorescent microplastics in fluorescence microscopy images using deep learning-based image segmentation to enable accurate measurement of microplastic contamination.

Key Features:

• Deep Learning-Based Segmentation: Uses deep learning to distinguish fluorescent microplastics from other elements in fluorescence microscopy images.
• Model Ensemble: Evaluates nine deep learning models, six of which are based on the U-Net architecture, with MP-Net derived from U-Net.
• Training Data: Trained on at least 20,000 image patches extracted from 99 fluorescence microscopy images of microplastics from clams paired with binary masks.
• Performance Metrics: Achieved a mean F1-score of 0.736 and a mean Intersection over Union (IoU) of 0.617, outperforming other evaluated models.
• Test-Time Augmentation: Applied brightness, contrast, and HSV adjustments at test time to evaluate robustness, without a clear improvement in predictive performance.
• Quantitative Recovery Assessments: Predicted microplastic quantities in spiked and real images that closely matched ground truth values.
• Integration with MAP: Integrated into the Microplastics Annotation Package (MAP) to automate MP quantification and interoperate with MP-VAT, while supporting manual annotation and model fine-tuning.

Scientific Applications:

• Environmental Monitoring: Quantifying fluorescent microplastics in microscopy images to assess contamination levels in environments and specimens such as clams.
• Public Health and Research: Providing quantitative data for studies on microplastic distribution and potential health risks associated with microplastic ingestion.

Methodology:

Variants of MP-Net (nine models, six U-Net–based) were trained on ≥20,000 patches from 99 fluorescence microscopy images with corresponding binary masks; model selection used mean F1-score and mean IoU (0.736 and 0.617); test-time augmentations (brightness, contrast, HSV) were applied; quantitative evaluation included recovery assessments on spiked and real images; MP-Net was integrated into MAP for automated quantification and interoperability with MP-VAT.