A Deep Learning Approach for Microplastic Segmentation in Microscopic Images

Area of Science:

• Environmental Science
• Ecotoxicology
• Computer Science

Background:

• Microplastic pollution is widespread, posing ecotoxicological risks.
• Current analytical methods for microplastic characterization are slow and inefficient.
• Microplastic morphology significantly influences toxicological effects, necessitating detailed analysis.

Purpose of the Study:

• To develop a novel deep learning model for high-throughput microplastic segmentation and morphological characterization.
• To address the measurement bottleneck in current microplastic analysis.
• To provide toxicologically-relevant data for accurate risk assessment.

Main Methods:

• Developed MNv4-Conv-M-fpn, a deep learning model for multi-class microplastic segmentation.
• Utilized transfer learning, a Feature Pyramid Network, and a Feature Fusion Module.
• Segmented microscopic images into six classes: five microplastic types (fiber, fragment, sphere, foam, film) and background.

Main Results:

• The MNv4-Conv-M-fpn model achieved high accuracy and computational efficiency.
• Demonstrated near real-time inference speed, outperforming existing segmentation methods.
• Validated performance on a diverse dataset, showing robustness and low computational load.

Conclusions:

• MNv4-Conv-M-fpn offers a precise and scalable solution for microplastic analysis.
• The model is suitable for high-throughput environmental monitoring and resource-constrained settings.
• Enables more accurate and efficient assessment of microplastic pollution in ecosystems.