The Active Segmentation Platform for Microscopic Image Classification and Segmentation.

Area of Science:

• Computer Vision
• Biomedical Image Analysis
• Machine Learning

Background:

• Traditional image segmentation methods are problem-specific and lack universal applicability.
• Machine learning offers pixel-level classification but struggles with appropriate feature selection.
• Automated solutions are needed to overcome limitations in current image segmentation and classification techniques.

Purpose of the Study:

• To present an automated image segmentation and classification platform, Active Segmentation, based on ImageJ.
• To integrate expert domain knowledge with geometrical feature extraction and machine learning.
• To demonstrate the platform's efficacy in both segmentation and whole image classification tasks.

Main Methods:

• Developed the Active Segmentation platform integrating ImageJ, expert knowledge, and partial ground truth.
• Employed geometrical feature extraction using multi-scale signal processing.
• Combined feature enrichment with feature selection for machine learning classifiers.

Main Results:

• Achieved performance comparable to deep learning approaches in image segmentation and classification.
• Demonstrated successful application on the ISBI 2012 segmentation challenge dataset.
• Validated whole image classification on HeLa and HEp-2 datasets.

Conclusions:

• Automated image segmentation and classification can be effectively achieved by integrating domain knowledge and advanced feature engineering.
• Differential geometry significantly enhances machine learning outcomes by enriching feature spaces with geometrical invariant objects.
• The Active Segmentation platform offers a robust and versatile solution for complex image analysis tasks.