Background

Segmentation of nuclei in whole slide images is a common methodology in pathology image analysis. A variety of anatomical segmentation methods have been developed to detect the location of nuclei and other cellular structures.  Achieving accurate and robust segmentation results remains a difficult problem because of image noise. Therefore it is necessary to have a quality control stage to assess the quality of segmentation results before the results are used in further analysis for discovery or interpretation.  In view of the vast amount of date contained within whole slide pathology images, universal automated error checking methods are needed to help pathologist detect bad segmentation quickly and reliably.

Technology

Researchers at Stony Brook University have developed a system for performing quality assessment of the segmentation of digital pathology images using a machine learning approach. This system is based on dividing the image into patch level texture features which are then the basis of the quality assessment. The machine learning classifier is trained using images that have been labeled by pathology experts and then partitioned into disjoint rectangular image patches. Once trained, the classification model can then be used to assess new images on a patch level, overlying a heat map as to which areas of the image of been segment adequately or poorly (Figure 1). Further Details : Wen S, Kurc TM, Gao Y, Zhao T, Saltz JH, Zhu W. A methodology for texture feature-based quality assessment in nucleus segmentation of histopathology image. J Pathol Inform 2017;8:38.

Advantages

- Scalable Process: minimized computing power for Q/A assessment of large data sets - Universal application across different segmentation algorithms - (Semi-)Automated

Application

- Digital pathology - Industrial applications include satellite and astronomical image data analysis.