Influence of normalization and color features on super-pixel classification: application to cytological image...

Area of Science:

• Medical Image Analysis
• Computational Biology
• Computer Vision

Background:

• Accurate super-pixel feature extraction is crucial for color super-pixel classification.
• Automatic recognition of nucleus and cytoplasm in cytological images presents a significant challenge.

Purpose of the Study:

• To propose and evaluate an automatic white blood cell (WBC) segmentation method using super-pixel classification.
• To determine optimal color normalization, color spaces, and feature extraction techniques for WBC segmentation.

Main Methods:

• The proposed method involves five steps: color normalization, super-pixel generation (Simple Linear Iterative Clustering), illumination invariance, color feature extraction, and supervised classification.
• An exhaustive statistical evaluation of various normalization methods, color spaces (including RGB), and feature extraction techniques was conducted.

Main Results:

• Comprehensive gray world normalization significantly improved super-pixel classification accuracy compared to no normalization, ranking highest in the Friedman test.
• The RGB color space was found to be optimal for super-pixel feature extraction in this context.
• First-order statistics features, combined with learning methods, performed best for automatic WBC segmentation.

Conclusions:

• The proposed super-pixel classification approach provides an effective method for automatic WBC segmentation.
• Comprehensive gray world normalization and the RGB color space are recommended for enhanced performance in nucleus and cytoplasm recognition.
• First-order statistics features are suitable for feature extraction in automated WBC segmentation tasks.