Segmentation of fluorescence microscopy cell images using unsupervised mining.

Area of Science:

• Medical Informatics
• Computational Biology
• Image Analysis

Background:

• Accurate cell and nuclei contour measurement is vital for detecting cellular changes in medical informatics.
• Microscopy with fluorescence staining uses image segmentation as a primary step.
• Unsupervised clustering methods can enhance cell segmentation, addressing microscopy complexities.

Purpose of the Study:

• To develop and evaluate unsupervised data mining techniques for cell image segmentation.
• To compare the performance of k-means, EM, Otsu's threshold, and GMAC algorithms.
• To assess the efficacy of these methods for complex real-world cell image analysis.

Main Methods:

• Adaptation and evaluation of four unsupervised learning methods: k-means clustering, Expectation-Maximization (EM), Otsu's thresholding, and Gaussian Mixture Autoregressive Clustering (GMAC).
• Quantitative and qualitative performance evaluation using synthetic and real microscopy data.
• Definition of validation measures to assess segmentation accuracy.

Main Results:

• K-means, Otsu's threshold, and GMAC demonstrated similar and precise segmentation results.
• The EM algorithm exhibited higher recall but lower precision due to under-segmentation, attributed to its Gaussian model assumption.
• Spatial information is crucial for effective segmentation of complex cell images.

Conclusions:

• K-means, Otsu's threshold, and GMAC are effective unsupervised methods for cell image segmentation.
• The EM algorithm's performance is limited by its Gaussian model assumption in this context.
• Integrating spatial information is recommended for robust cell segmentation in medical informatics.