Weakly Supervised Collaborative Learning for Airborne Pollen Segmentation and Classification from SEM Images.

Area of Science:

• Microscopy and Imaging
• Computational Biology
• Machine Learning

Background:

• Pollen identification accuracy is limited by image quality and impurities in Scanning Electron Microscopy (SEM) images.
• Existing methods struggle with real-world SEM data containing various contaminants.
• Weakly supervised learning offers a potential solution for pollen analysis with limited labeled datasets.

Purpose of the Study:

• To develop a collaborative learning framework for joint pollen segmentation and classification.
• To improve pollen identification accuracy in the presence of impurities in SEM images.
• To achieve high performance with limited image-level supervision.

Main Methods:

• A detection model is used to locate pollen regions in raw SEM images.
• A pre-trained U-Net model segments pollen grains using unsupervised contour features.
• Deep convolutional neural networks refine segmentation masks using activation maps for collaborative training.

Main Results:

• The proposed method effectively mitigates interference from impurities in SEM images.
• Pollen identification accuracy reached 86.6% using limited supervision (approx. 1000 images).
• Collaborative training improved both segmentation and classification performance.

Conclusions:

• The collaborative learning approach enhances pollen identification accuracy and robustness against impurities.
• This method demonstrates the effectiveness of weakly supervised learning for pollen analysis.
• The technique offers a viable solution for pollen identification with limited labeled SEM data.