ArtSeg-Artifact segmentation and removal in brightfield cell microscopy images without manual pixel-level annotations

Area of Science:

• Life Sciences
• Biotechnology
• Medical Imaging

Background:

• Brightfield microscopy is crucial in life sciences but images often contain artifacts.
• These artifacts impede accurate downstream analyses like cell segmentation and quantification.
• Current artifact removal methods often require laborious pixel-level annotations.

Purpose of the Study:

• To introduce ScoreCAM-U-Net, a novel deep learning pipeline for automated artifact segmentation in brightfield microscopy images.
• To enable efficient artifact removal with minimal user input, utilizing image-level labels instead of pixel-level annotations.
• To demonstrate the effectiveness of automated artifact removal in improving the reliability of subsequent image analyses.

Main Methods:

• Development of the ScoreCAM-U-Net architecture, a deep convolutional neural network.
• Training the model using image-level labels for artifact presence/absence, significantly reducing annotation time.
• Validation of the model on three distinct brightfield microscopy datasets with diverse artifact types.

Main Results:

• ScoreCAM-U-Net successfully segments artifactual regions in brightfield images with high performance.
• The model's training process is orders of magnitude faster than traditional pixel-level annotation methods.
• Automated artifact removal demonstrably improved accuracy in nuclei segmentation, morphometry, and fluorescence intensity quantification.

Conclusions:

• ScoreCAM-U-Net offers a rapid and effective solution for artifact removal in brightfield microscopy.
• The use of image-level labels makes deep learning-based artifact segmentation more accessible and scalable.
• This automated approach is poised to become an essential step in large-scale microscopy experiments for enhanced data integrity.