Bayesian level sets and texture models for image segmentation and classification with application to non-invasive stem cell monitoring

Abstract

Image segmentation and classification, the identification and demarcation of regions of interest within an image, is necessary prior to subsequent information extraction, analysis, and inference. Many available segmentation algorithms require manual delineation of initial conditions to achieve acceptable performance, even in cases with high signal to noise and interference ratio that do not necessitate restoration. This weakness impedes application of image analysis to many important fields, such as automated, mass scale cultivation and non-invasive, non-destructive high throughput analysis, monitoring, and screening of pluripotent and differentiated stem cells, whether human embryonic (hESC), induced pluripotent (iPSC), or animal. Motivated by this and other applications, the Bayesian Level Set (BLS) algorithm is developed for automated segmentation and classification that computes smooth, regular segmenting contours in a manner similar to level sets while possessing a simple, probabilistic implementation similar to that of the finite mixture model EM. The BLS is subsequently extended to harness the power of image texture methods by incorporating learned sets of class-specific textural primitives, known as textons, within a three-stage Markov model. The resulting algorithm accurately and automatically classifies and segments images of pluripotent hESC and trophectoderm colonies with 97% and 91% accuracy for high-content screening applications and requires no previous human initialization. While no prior knowledge of colony class is assumed, the framework allows for its incorporation. The BLS is also validated on other applications, including brain MRI, retinal lesions, and wildlife images.