A methodology for quality control in cell nucleus segmentation

In order to achieve the very high accuracy rates required in unsupervised automated biomedical applications, it is often necessary to complement a successful segmentation algorithm with a robust error checking stage. The better the segmentation strategy, the less severe the error checking decisions need to be and the fewer correct segmentations that are discarded. These issues are dealt with in this paper in order to achieve 100% accuracy on a data set of 19946 cell nucleus images using an established segmentation scheme with a success rate of 99.47%. The method is based upon measuring changes in the final segmentation contour as the one parameter that governs its behaviour is varied. 1. Introduction remove potential artefacts based on shape and appearance that was capable of detecting some of the incorrectly segmented nuclei [9]. Nordin describes an algorithm that is able to report a failure at various levels of segmentation, as well as a separate artefact rejection stage [11]. McKenna used a neural network to preselect potential nuclei in scenes for subsequent segmentation. It was pointed out that a post-processing stage would also be necessary to filter out 'erroneously detected objects'[10]. A common trait in these techniques is the use of a separate process to view the output of the segmentation and to use shape and appearance measurements to classify the results as 'pass' (looks like a cell) or 'fail' (doesn't look like a cell). We have proposed a segmentation scheme that not only employs an algorithm with much better performance than previously reported [3], but also enables a confidence measure in the resulting segmentation to be given.