A Novel Performance Evaluation System for Fluorescent Cell Image Segmentation

Image segmentation is crucial to modern cell biology which depends on quantitative analysis of fluorescent microscopy images. Segmented regions are useful to estimate localization and dynamics of cells and sub-cellular objects. Although many segmentation algorithms have been proposed in image processing and pattern recognition fields, most approaches are designed for some specific tasks and may not work well for other tasks. This makes it difficult to find an algorithm suitable for a given task for biologists who do not have knowledge about segmentation algorithms. Automatic selection of an appropriate segmentation algorithm including its parameters efficiently has become a very important duty, since recent advances in cell and sub-cellular imaging technology significantly increases a number of observed images for biologists. In this paper, we propose a novel segmentation system which has the function of performance evaluation. The set of candidate algorithms including parameter variation are automatically generated by using combinations of the prescribed image processing methods implemented in our system, and the system is designed in order to add and replace these algorithms easily. Their segmentation qualities are evaluated by comparing each result with the ground-truth provided by biologists based on either a single or multiple similarity metrics. Finally, our system predicts which algorithm will provide the best performance on a set of images similar to the original image with ground-truth reference. We examine our system using typical segmentation algorithms under several evaluation metrics and find it useful especially for detection of fluorescent labeled targets with granular shapes on real sub-cellular images, as well as simulated images with small sub-cellular objects.