Practical guidelines for cell segmentation models under optical aberrations in microscopy.

Area of Science:

• Biomedical imaging
• Computational biology
• Microscopy

Background:

• Cell segmentation is crucial for analyzing cellular morphology and behavior in biomedical research.
• Deep learning, especially Convolutional Neural Networks (CNNs), excels at cell segmentation but struggles with optical aberrations.
• Evaluating model robustness under various aberrations is vital for reliable biological image analysis.

Purpose of the Study:

• To assess the performance of cell image segmentation models under simulated optical aberrations.
• To compare the robustness of different deep learning models and traditional methods on fluorescence and bright field microscopy datasets.
• To introduce a novel model for identifying aberration types and amplitudes to guide segmentation model selection.

Main Methods:

• Simulated various optical aberrations (astigmatism, coma, spherical aberration, trefoil, mixed).
• Evaluated Otsu threshold, Mask R-CNN (with FPN/C3 heads, ResNet/VGG/Swin Transformer backbones), and Cellpose 2.0 on DynamicNuclearNet and LIVECell datasets.
• Developed and validated the Point Spread Function Image Label Classification Model (PLCM) for aberration characterization.

Main Results:

• The FPN with SwinS backbone combination showed superior robustness for simple cell images with minor aberrations.
• Cellpose 2.0 demonstrated effectiveness for complex cell images under similar aberrated conditions.
• PLCM accurately identified aberration types and amplitudes from Point Spread Function (PSF) images.

Conclusions:

• Specific deep learning architectures offer varying degrees of robustness against optical aberrations in cell segmentation.
• Cellpose 2.0 is recommended for complex cell images with aberrations.
• PLCM aids researchers in selecting appropriate segmentation strategies by characterizing optical aberrations, improving the reliability of cell analysis.