A stochastic reshapable contour model for microscopic image segmentation

Abstract

Active contours and active shape models have been widely employed in bio- medical image segmentation. However, their performance can be significantly decreased by two major reasons. 1) The image exhibit sophisticated spatial color patterns. 2) There is occlusion between interested objects of the image. We pro- posed a novel stochastic reshapable contour model to overcome these limitations of classic active contours. The cost functional of the proposed model comprised of three terms including the prior shape and regional texture information as well as the gradient information. The suggested model has the following practical advantages: 1) ability to be tuned and applied to wide range of bio-medical images; 2) robustness to occlusion and clutter; 3) ability to segment multiple objects of interest in the same image. Experimentation was carried out using a heterogeneous histology image data set which contains 462 manually-segmented cells. The segmentation performance of the proposed model is compared with manually segmented ground truth. The assessments demonstrate consistent and generally acceptable segmentation performance of the model.