Segmentation of cellular ultrastructures on sparsely labeled 3D electron microscopy images using deep learning.

Area of Science:

• Electron microscopy and deep learning for cancer research.

Background:

• Focused ion beam-scanning electron microscopy (FIB-SEM) offers detailed cellular ultrastructure of tumor cells.
• Analyzing these structures is crucial for understanding cancer mechanisms but is hindered by segmentation challenges.

Purpose of the Study:

• To develop and validate a deep learning approach for segmenting cells and subcellular ultrastructures in 3D FIB-SEM images of metastatic breast and pancreatic tumors.
• To overcome the bottleneck in quantitative analysis of tumor ultrastructure.

Main Methods:

• Utilized deep learning (neural networks) for segmenting subcellular organelles (nuclei, nucleoli, mitochondria) with high accuracy.
• Developed a multi-pronged approach for cell segmentation combining intracellular space detection, optical flow for boundary propagation, and filopodia tracking.
• Trained neural networks with sparse manual labels for ultrastructure segmentation.

Main Results:

• Achieved high Dice scores for segmentation: 0.93 for cells, 0.99 for nuclei, 0.98 for nucleoli, and 0.86 for mitochondria.
• Demonstrated accurate segmentation of relatively well-defined ultrastructures and challenging cell boundaries.
• The method enables quantitative image features to be associated with clinical variables.

Conclusions:

• Deep learning-based segmentation significantly improves the analysis of 3D FIB-SEM tumor images.
• This approach facilitates detailed quantitative analysis of cellular and subcellular organization in cancer.
• The findings pave the way for enhanced interpretative rendering and clinical correlation of ultrastructural data.