Comparing ensemble methods combined with different aggregating models using micrograph cell segmentation as an...

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Deep neural networks (DNNs) can suffer from high variance, impacting performance in tasks like cell segmentation.
• Ensemble learning and averaging techniques aim to improve the robustness and accuracy of DNNs by combining multiple models.

Purpose of the Study:

• To evaluate various ensemble averaging techniques for micrograph cell segmentation using deep convolutional neural networks (DCNNs).
• To compare the effectiveness of different ensemble formation strategies and averaging methods on segmentation performance.

Main Methods:

• Ensembles were formed using random seeds, L1-norm pruning, varying training data, or combinations thereof.
• Averaging methods evaluated included mean, median, alpha-stable distribution location parameter, and majority vote of class membership probabilities (CMPs).
• Performance was assessed using accuracy and Dice coefficient on a cell segmentation task with a VGG-like DCNN architecture.

Main Results:

• The choice of ensemble averaging method had a marginal influence on segmentation accuracy and Dice coefficient.
• Simple mean averaging proved highly competitive compared to more complex methods like fitting alpha-stable distributions to CMPs.
• The ensemble formation strategy also showed limited impact on the final performance metrics.

Conclusions:

• For micrograph cell segmentation, simple mean ensemble averaging is a practical and effective strategy.
• Complex averaging methods offer little performance benefit over simple mean averaging in this specific application.
• The findings suggest that focusing on robust ensemble formation and straightforward averaging is sufficient for achieving high performance in cell segmentation tasks.