Comparative Study of Deep Transfer Learning Models for Semantic Segmentation of Human Mesenchymal Stem Cell...

Area of Science:

• Biomedical Engineering
• Cell Biology
• Computer Science

Background:

• Accurate segmentation of human mesenchymal stem cells (MSCs) is crucial for cell biology research.
• Manual analysis of cell micrographs is time-consuming and prone to variability.
• Automated segmentation methods using deep learning can improve efficiency and consistency.

Purpose of the Study:

• To comparatively evaluate the performance of U-Net, DeepLabV3+, SegNet, and Mask R-CNN for semantic segmentation of MSC micrographs.
• To identify the most effective deep learning model for automating MSC image analysis.

Main Methods:

• A dataset of 320 expert-annotated human MSC micrographs was utilized.
• Four semantic segmentation models (U-Net, DeepLabV3+, SegNet, Mask R-CNN) were trained using transfer learning with ImageNet pre-trained weights.
• Model performance was assessed using Dice coefficient and Jaccard index metrics.

Main Results:

• U-Net demonstrated superior segmentation accuracy with a Dice coefficient of 0.876 and a Jaccard index of 0.781.
• DeepLabV3+ and Mask R-CNN also achieved high performance, outperforming SegNet.
• SegNet yielded the least accurate segmentation results among the evaluated models.

Conclusions:

• The U-Net model is highly effective and recommended for automated semantic segmentation of human mesenchymal stem cell micrographs.
• Implementing U-Net can significantly streamline routine cell culture analysis in biomedical laboratories.
• This automation facilitates more efficient and reliable research in cell biology.