Benchmarking of deep learning algorithms for 3D instance segmentation of confocal image datasets.

Area of Science:

• Biomedical imaging
• Cellular biology
• Computational microscopy

Background:

• Accurate segmentation of 3D microscopy images is crucial for understanding cellular processes.
• Deep learning (DL) methods are emerging as state-of-the-art for image segmentation.
• Lack of standardized evaluation hinders comparison of DL segmentation pipeline performance.

Purpose of the Study:

• To inventory and compare the performance of various DL pipelines for 3D cell segmentation.
• To evaluate DL methods against a non-DL method (MARS) using a common dataset.
• To assess segmentation accuracy under various image artifacts and establish a clear evaluation strategy.

Main Methods:

• Implemented and quantitatively compared several representative DL pipelines for 3D cell segmentation.
• Trained DL models on a common dataset of 3D cellular confocal microscopy images.
• Utilized a specific evaluation method to distinguish under- and oversegmentation errors, complemented by 3D visualization.

Main Results:

• DL pipelines exhibit varying segmentation accuracies.
• Two end-to-end 3D DL methods, originally for cell boundary detection, demonstrated superior performance.
• These high-performing DL methods showed significant adaptability to new datasets.

Conclusions:

• Not all DL pipelines are equal for 3D cell segmentation; performance varies significantly.
• End-to-end 3D DL approaches for cell boundary detection offer a promising direction for accurate and adaptable cellular image segmentation.
• Standardized evaluation and visualization are key to understanding and advancing DL in this field.