Bayesian inference for the automated adjustment of an image segmentation pipeline — A modular approach applied to wound healing assays

Abstract

Background:Dynamical biological and physiological processes as the migration of single cells, collective cell movement during tissue formation or the metastasis of tumors can nowadays be observed under in-vitro and in-vivo conditions. High temporal and spatial resolution require automated image segmentation and analysis. Although, open source and commercial software allow the segmentation of target regions, all parameters of an appropriate image processing algorithm have to be adapted manually by the user. Typically the experimenter knows details about the resulting images whereas he might not be trained to adapt parameters of segmentation algorithms. Methods:It is the aim of this study to provide an automated estimation of these parameters by applying Bayesian data analysis based on a single manually segmented image for calibration. We apply this technique to a temporal sequence of images showing the closing of a wound. The corresponding likelihood is given as difference between the manually segmented contour of the wound and the resulting model boundary of the segmentation process. We apply a typical segmentation pipeline consisting of an edge filter, a blurring filter and an area cut off process where six parameters control these operations. Bayesian multinested sampling algorithm is applied to estimate automatically these image pipeline parameters and their uncertainties. Results:The proposed algorithm is logically consistent and performs image segmentation with a high level of accuracy especially with regard to inter-observer variability in the input data. Further, Bayesian data analysis allows to estimate the uncertainty of the segmented wound area and of the velocity of the closing boundary. Conclusion:We were able to introduce a new approach for automated image segmentation, which produces excellent results in terms of ease of handling, preservation of expert knowledge, robustness and displaying its own uncertainties. Due to a broadly modular approach, the presented technique can also be applied to other processing pipelines offering a pragmatic and robust way to obtain an automated segmentation of biomedical data driven by the prior knowledge and information specified by the experimenter.