Neural network fast-classifies biological images through features selecting to power automated microscopy

Area of Science:

• Computational Biology
• Microscopy
• Artificial Intelligence

Background:

• Current optical microscopy relies on post-acquisition analysis for cell detection and classification.
• Future smart microscopy requires on-the-fly image analysis for automated acquisition decisions.
• Biological sample preparation and expert annotation are time-consuming and costly, limiting dataset size.

Purpose of the Study:

• To propose a real-time image processing method for smart microscopy.
• To balance accurate cell detection and classification with high execution performance.
• To optimize feature selection for efficient machine learning analysis in microscopy.

Main Methods:

• Characterized images using a generic, high-dimensional feature extractor.
• Employed machine learning classifiers (Random Forests, Fisher's linear discriminant) to analyze feature contributions.
• Developed a method to select fast and discriminant features, excluding redundant ones.

Main Results:

• Random Forests outperformed Fisher's linear discriminant for cell classification.
• Excluding time-consuming and less discriminant features significantly reduced execution time without substantial accuracy loss.
• Achieved 79.6% accuracy in classifying cells into eight cell cycle phases within 68.7 ms per cell on an embedded system.

Conclusions:

• A strategy for selecting fast and discriminant features enables efficient real-time image analysis for smart microscopy.
• The proposed method significantly improves execution speed, allowing for 14 cells per second classification.
• This approach is adaptable for neural networks and GPUs, paving the way for optimized smart microscopy algorithms.