Segmentation of mitochondria in fluorescence micrographs by SVM

Mitochondrial morphology is correlated with mitochondrial functions, neurodegenerative diseases and ageing. Mitochondrial morphology may also be important to study the etiology of type 2 diabetes due to vital role of β-cell mitochondria in insulin. We have developed a quantification system for numerical analysis of mitochondrial morphology of β-cell. However, accuracy of our previous system for certain type of image is only ~60% due to errors in segmentation of mitochondria. A new segmentation must be developed to improve our analysis system. Support Vector Machine (SVM) is considered as good approach candidate for better segmentation because this method separates the classes in certain way to maximize the margin among them. Five manually segmented single cell mitochondria images of five typical morphological subtypes are used as training data. The new SVM-based segmentation method shows its better performance by comparing with three traditional methods, including top-hat, Mean and Otsu. The single cell fluorescence micrographs are used as test data for checking performance of our new system. Our new system can segment mitochondria accurately even mitochondria with various intensities are crowded in noisy background.