Deep learning based microscopic cell images classification framework using multi-level ensemble.

Area of Science:

• Computational Biology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated artificial intelligence (AI)-based microscopic image classification systems are advancing rapidly.
• Existing systems often classify specific cell types from limited microscopy techniques.
• A need exists for a generic framework applicable to diverse cell types and microscopy methods.

Purpose of the Study:

• To develop a generic framework for classifying microscopic cell images of multiple types.
• To enable classification across a variety of microscopy techniques.
• To enhance the general applicability of AI in cell image analysis.

Main Methods:

• A transfer learning-based multi-level ensemble approach was developed.
• The ensemble trained base models using varied optimization methods and learning rates.
• Batch normalization was incorporated for faster convergence in the ensemble models.

Main Results:

• The framework demonstrated general applicability across five public datasets.
• The proposed method achieved superior performance compared to existing similar works.
• The system effectively captured multi-scale features from microscopic cell images.

Conclusions:

• The developed framework outperforms state-of-the-art methods for microscopic cell classification.
• The approach requires comparatively less training data for effective classification.
• This generic framework advances automated analysis in microscopic cell imaging.