Multi-Modality Microscopy Image Style Augmentation for Nuclei Segmentation.

Area of Science:

• Medical image analysis
• Computational biology
• Machine learning

Background:

• Manual annotation of microscopy images for nuclei segmentation is labor-intensive and time-consuming.
• Existing annotation datasets are often limited in size and diversity across different imaging modalities and experimental conditions.

Purpose of the Study:

• To develop a novel data augmentation technique for microscopy images using generative adversarial networks (GANs).
• To improve the robustness and accuracy of nuclei segmentation algorithms by addressing data heterogeneity and class imbalance.

Main Methods:

• A microscopy-style augmentation technique based on a generative adversarial network (GAN) was proposed.
• The GAN utilizes disentangled representations for content and style to preserve image structure while altering style.
• The method was evaluated on the 2018 Data Science Bowl dataset, encompassing various cell assays, lighting conditions, and imaging modalities (bright-field and fluorescence).

Main Results:

• The proposed style augmentation significantly increased the segmentation accuracy of top-ranked Mask R-CNN-based nuclei segmentation algorithms.
• The technique demonstrated effectiveness in handling diverse cell assay types, lighting conditions, and imaging modalities.
• The augmentation method improved robustness to test data heterogeneity and addressed class imbalance without resampling.

Conclusions:

• The novel GAN-based style augmentation technique enhances nuclei segmentation performance in microscopy images.
• This approach effectively addresses data heterogeneity and class imbalance, leading to more robust downstream segmentation tasks.
• The method offers a valuable tool for leveraging limited annotated data across multiple microscopy modalities.