Computer-aided detection of rare tumor populations in flow cytometry: an example with classic Hodgkin lymphoma

Diagnosing classical Hodgkin lymphoma (cHL) by flow cytometry (FC) relies on an observer gating rare populations of Hodgkin/Reed Sternberg (HRS) cells. Here, we apply machine-learning methods to aid in the detection of rare tumor cell populations using data derived from clinical FC analysis of cHL as a model disease.

FC data from 144 clinical cases using a nine-color FC reagent panel were analyzed using Python 2.7 and the "scikit-learn" module.

Seventy-eight 50 × 50 two-dimensional histograms were generated from routine FC data and a reciprocal power function applied to favor rare events. Data were classified by support vector machine (SVM), gradient boosting, and random forest classifiers. All three classifiers showed no statistical difference in performance, with 89%-92% accuracy on cross-validation. Nearly all classifiers misclassified the same set of cases, with more false-positive than false-negative cases. Dimensionality reduction by ensemble methods selected for data points in a CD5+/ CD40+/CD64- region.

All classifiers provide probabilistic confidences for each result, and diagnostic cutoffs can be chosen to minimize false negatives and serve as a screening tool. Computational exclusion of manually gated HRS cells had little impact on the overall performance of selected support vectors in SVM or dimensionality reduction, suggesting that features of the immune response in cHL may dictate the method accuracy. We hypothesize there are distinct inflammatory cells that suggest cHL.