RESEARCH ABSTRACT

Quantitative scoring method to assess computer-based methods of histological stain normalization and generation of high quality cancer pathology image data sets suitable for machine learning applications.

​

The lack of standardized pathology images limits machine learning advancements in the field of digital cancer pathology.  Traditional cancer pathology involves evaluation of tissue samples to determine if tissue is cancerous and, if so, to assess cancer invasiveness, grade, mitotic rate and other factors.  Today, this process depends on qualitative (human eye) review.  A superior approach incorporating digital pathology would supplement human analysis with quantitatively-derived insights, such as those formulated using machine learning algorithms that compare patient images with “positive” and “negative” diagnostic images. However, high quality pathology image data sets, particularly of Whole Slide Images (WSIs), have not been adequately standardized to provide training datasets for machine learning.  Laboratory inconsistencies in the tissue staining process can cause variations in image exposure and color.  Digital post-processing of images through methods such as color normalization can improve image consistency. However, current techniques rely on qualitative assessment to measure the success of color normalization techniques. There is also no universal approach to comparing the results of different color normalization techniques.  These are significant obstacles to the creation of the high quality standardized image libraries needed for machine learning.  This study devised a method of proof to validate the success of the color normalization methods tested and proposed a means to quantitatively measure the effectiveness of such color normalization methods, known as the Color Correspondence Score.  In addition, this study successfully demonstrated the viability of using low cost and open source software to perform the computationally expensive task of normalizing gigapixel resolution WSIs.

​

​