Impact of Training Data, Ground Truth and Shape Variability in the Deep Learning-Based Semantic Segmentation of HeLa...

Area of Science:

• * Biomedical image analysis
• * Deep learning in microscopy

Background:

• * Accurate cell segmentation is crucial for biological research.
• * Deep learning models like U-Net show promise for image segmentation.
• * Evaluating the impact of training data size and quality is essential for model performance.

Purpose of the Study:

• * To investigate how training data amount and variability affect U-Net segmentation of HeLa cells.
• * To assess the correctness of manually generated ground truth data.
• * To compare U-Net performance against traditional image processing algorithms.

Main Methods:

• * Used 3D electron microscopy images of HeLa cells (8192×8192×517).
• * Cropped and manually delineated a region of interest (ROI) for ground truth.
• * Generated data/label pairs for nucleus, nuclear envelope, cell, and background.
• * Trained U-Net architectures with varying numbers of data pairs (36,000 to 270,000).
• * Compared results from manually segmented and automatically generated data.

Main Results:

• * Increased training data pairs led to improved accuracy and Jaccard similarity index for ROI segmentation.
• * U-Net trained with automatically generated data outperformed manual segmentation on larger slices.
• * Combining manually and automatically generated data (270,000 pairs) yielded the best segmentation performance.

Conclusions:

• * The quantity and diversity of training data significantly impact U-Net segmentation performance.
• * Automatically generated training data can effectively represent cell variability for improved segmentation.
• * Optimizing training data generation and quantity is key for robust deep learning-based cell image analysis.