|
Circulating tumor cell clusters (CTCCs) are associated with high metastatic potential and poor patient prognosis. However, they are difficult to detect and isolate because of their extremely low numbers. Here, we report on the use of machine learning based analysis to achieve highly accurate detection of CTCCs in flowing whole blood samples relying on the confocal detection of endogenous light scattering and fluorescence signals. Our custom flow cytometer utilizes laser excitation at 405, 488, and 633 nm and confocal detection of the corresponding light scattering signals as well as fluorescence in the 525 25nm and 67020nm range. Samples consist of whole blood isolated from mice or rats spiked with varying concentrations of CTCCs consisting of 2-15 cell CTCCs, flowed through the channels of a microfluidic device. The CTCCs utilized in this initial study express GFP, so that we can detect the strong GFP signal using the 525 nm detector and use that signal as the ground truth for assessing the performance of algorithms relying on the endogenous signals detected by the other detectors. Our data is acquired during 18 independent experiments with data from 13 days used for training and five days used for testing. There are over 6,000 true positive and over 60,000 false negative peaks in this data set. We consider narrow neural network, fine k-nearest neighbors and ensemble bagged tree (EBT) models and we find that an EBT with gentle boost model yields optimal performance. Using the data from the three light scattering channels and the autofluorescence channel results in CTCC detection with purity, sensitivity, specificity and accuracy that exceed 90% in the test data. These promising results motivate further development of label-free flow cytometry using non-GFP expressing CTCCs for in vitro and in vivo applications.
|
