Distinguishing Reactive Lymphocytes From Blasts Using Fractal Chromatin Patterns.

Abstract

Introduction: Of all the cells identified in peripheral blood smears, reactive lymphocytes (RLs) and blasts are considered especially difficult to differentiate. Blasts and RLs are present in distinct diseases that carry unique prognoses and treatments; however, there are currently no definitive methods to distinguish these cells morphologically.

Methods: We developed a method to distinguish between blasts and RLs based on the quantification of fractal chromatin patterns. Nuclei from white blood cell images were isolated, and the fractal patterns were quantified using The Workflow of Matrix Biology Informatics (TWOMBLI) software. Quantified fractals were compared using the t-test. The data was further split into training and testing sets. Models (random forest and k-nearest neighbors) were selected through cross-validation on the training sets. Performance metrics, including area under the curve (AUC), accuracy, precision, specificity, and sensitivity, were determined for the selected models on the testing sets. Principal component analysis (PCA) was also performed.

Results: Our most general model was able to identify RLs and blast subtypes with an average 84.2% accuracy and an AUC of 0.844. Testing on the holdout set gave every model an area under the curve greater than 0.815. PCA revealed two components that account for 50% of the data's variance.

Conclusion: Our results suggest that a classification algorithm can effectively distinguish between blasts and RLs based solely on fractal chromatin patterns. It is possible that a similar algorithm could be utilized in the clinical hematology laboratory to assist in distinguishing RLs and blasts in peripheral blood smears.