Memory-efficient semantic segmentation of large microscopy images using graph-based neural networks.

Area of Science:

• Computational Biology
• Image Analysis
• Machine Learning

Background:

• Deep learning models like CNNs are standard for image segmentation but struggle with large datasets due to memory constraints.
• Automating segmentation of large-scale microscopy images is crucial for biological research.

Purpose of the Study:

• To introduce a graph neural network (GNN)-based framework for large-scale microscopy image segmentation.
• To evaluate the performance of GNNs against CNNs in terms of accuracy, training time, and memory usage.

Main Methods:

• Images are converted into graphs using superpixels, enabling processing of large images with limited memory.
• A GNN framework is developed to perform segmentation on these graph representations.
• Performance is benchmarked against traditional CNN-based methods using microscopy images of cells.

Main Results:

• GNN-based segmentation utilized one to three orders of magnitude fewer computational resources than CNNs.
• Accuracy changes were minimal, ranging from -2% to +0.3%.
• The GNN framework's accuracy can be improved by optimizing superpixel generation.

Conclusions:

• GNNs provide a computationally efficient alternative for large-scale microscopy image segmentation.
• The framework offers a favorable trade-off between computational cost and accuracy compared to CNNs.
• This GNN approach is highly suitable for memory-limited biological image analysis tasks.