3D Microscopic Images Segmenter Modeling by Applying Two-Stage Optimization to an Ensemble of Segmentation Methods Using a Genetic Algorithm

Abstract

This paper presents research related to segmentation based on supervisory control, at multiple levels, of optimization of parameters of segmentation methods, and adjustment of 3D microscopic images, with the aim of creating a more efficient segmentation approach. The challenge is how to improve the segmentation of 3D microscopic images using known segmentation methods, but without losing processing speed. In the first phase of this research, a model was developed based on an ensemble of 11 segmentation methods whose parameters were optimized using genetic algorithms (GA). Optimization of the ensemble of segmentation methods using GA produces a set of segmenters that are further evaluated using a two-stage voting system, with the aim of finding the best segmenter configuration according to multiple criteria. In the second phase of this research, the final segmenter model is developed as a result of two-level optimization. The best obtained segmenter does not affect the speed of image processing in the exploitation process as its operating speed is practically equal to the processing speed of the basic segmentation method. Objective selection and fine-tuning of the segmenter was done using multiple segmentation methods. Each of these methods has been subject to an intensive process of a significant number of two-stage optimization cycles. The metric has been specifically created for objective analysis of segmenter performance and was used as a fitness function during GA optimization and result validation. Compared to the expert manual segmentation, segmenter score is 99.73% according to the best mean segmenter principle (average segmentation score for each 3D slice image with respect to the entire sample set). Segmenter score is 99.49% according to the most stable segmenter principle (average segmentation score for each 3D slice image with respect to the entire sample set and considering the reference image classes MGTI median, MGTI voter and GGTI).