Automated Classification of Normal and Leukemic B Cells on the ImageStream^® Imaging Flow Cytometer Using Immunocytomorphological Image Characteristics

Leukemic B cells can be distinguished from normal B cells based on their surface staining patterns and their morphology, but the differences can be subtle and the abnormal events rare. In order to reliably differentiate normal from leukemic B cells, objective quantitation of large data sets is required. This is accomplished with the ImageStream cytometer, which automatically acquires multispectral images of thousands of individual cells in flow at very high speeds. The associated IDEAS data analysis software measures hundreds of photometric and morphometric features per cell, thereby enabling simultaneous immunophenotyping and morphology-based measurements. The features quantify individual image characteristics such as shape, size, signal strength and texture. We have leveraged these features by implementing an automated classification routine that identifies the feature combination that best separates different cell types of interest. Our automated classification routine utilizes a modification of Fisher’s Linear Discriminant criterion to measure discrimination power between two or more populations of cells. The classifier generates a composite feature consisting of a weighted linear combination of basic image features that provides maximum discrimination. Thus, our method transforms a multi-parameter classification task that can be difficult to conceptualize into a classification based on a single composite feature that can be readily conceptualized and visualized using a simple histogram output. Since the composite feature is generated from, and tested and optimized on, large numbers of cells, we obtain results that are highly objective, repeatable, statistically significant and scalable with data size. We used this classification approach to study normal and tumor cells of the hematopoietic lineage with the goal of improving diagnosis of hematological disorders. Here we show the results of using automated classification to distinguish the subtle differences between B cells of CLL patients and those of normal patients. We also used the classifier to distinguish B-cell ALL from B-cell CLL. In both cases, we used a combination of shape, size, texture and signal strength features from the SSC, brightfield and nuclear imagery generated by the ImageStream system. The data presented demonstrate the application of a novel automated cell classification method to the discrimination of normal and tumor cells within large and heterogeneous cell samples.