Assessing microscope image focus quality with deep learning.

Area of Science:

• Microscopy image analysis
• Computational biology
• Machine learning for image processing

Background:

• Automated microscopy generates large datasets, often containing out-of-focus images that compromise data quality.
• Accurate identification of low-quality images is crucial for unbiased analysis.
• Previous methods struggled to provide absolute focus measures, relying on relative comparisons.

Purpose of the Study:

• To develop a deep neural network for accurate, absolute focus prediction in single microscope images.
• To enable interpretable predictions with patch-level analysis and certainty measures.
• To eliminate the need for manual annotation by using synthetic data.

Main Methods:

• A deep neural network was trained on synthetically defocused Hoechst stain images.
• The model predicts an absolute focus measure at the image-patch level.
• Prediction certainty is outputted for interpretability.

Main Results:

• The model achieved 95% accuracy in identifying absolute image focus within one defocus level.
• It outperformed previous methods on Hoechst and Phalloidin stain images for binary classification.
• The model demonstrated generalization to unseen stains (Tubulin) and cell types (MCF-7) on different instruments.

Conclusions:

• The deep neural network offers higher accuracy and precision in classifying out-of-focus images.
• Synthetic data training obviates the need for manual annotation, simplifying the process.
• The model's generalizability and availability as a free software library facilitate its adoption.