Cell Structure Segmentation in TEM Images of Murine Skin Melanoma Cells by Deep Learning Model

Area of Science:

• Cell Biology
• Biophysics
• Computational Biology

Background:

• Mitochondria-endoplasmic reticulum contact sites (MERCs) are vital for cellular functions including calcium homeostasis, lipid transport, and cell death.
• Understanding MERC dynamics is crucial for cancer therapeutics, yet their role in carcinogenesis is largely unknown.
• Current analysis of MERCs using transmission electron microscopy (TEM) is time-consuming, requiring manual processing of numerous images.

Purpose of the Study:

• To develop and evaluate deep learning models for automated segmentation of mitochondria, endoplasmic reticulum (ER), and MERCs from TEM images.
• To create a user-friendly tool for accelerating MERC analysis in tumor cells.
• To investigate the impact of pre-training on model performance for MERC segmentation.

Main Methods:

• Five U-Net models with ResNet34 encoders, including variations with attention and multi-level structure blocks, were trained and tested for image segmentation.
• Model performance was evaluated using F1 and Intersection over Union (IoU) metrics.
• The best-performing model, U-Net-scSE, was integrated into the free UltraNet web server for automated analysis.

Main Results:

• The U-Net-scSE model achieved the highest performance, with F1 scores of 0.872 for mitochondria and 0.744 for ER.
• Pre-training on external datasets did not significantly improve model performance on the test dataset.
• The developed UltraNet web server provides a free and user-friendly platform for automated MERC analysis from TEM images.

Conclusions:

• Deep learning, specifically the U-Net-scSE architecture, offers a powerful approach for accurate and efficient segmentation of cellular structures involved in MERCs.
• The UltraNet web server significantly accelerates the analysis of MERCs, facilitating further research into their role in cancer.
• Automated analysis tools are essential for advancing the study of complex cellular interactions like MERCs in the context of disease.