A supporting vector machine using morphological features of cells for classification of circulating tumor cell clusters

Abstract

Metastatic cancer is known to be caused by circulating tumor cells(CTCs), that are detached from the primary tumor and circulate in the bloodstream, and is the leading cause of death worldwide. Thus, the identification of the CTC and their clusters in the blood is crucial for predicting the risk of metastatic cancer and prognosis. Typically, CTCs are identified using immunofluorescent staining to differentiate them from the various cells in the blood, but the heterogeneity of CTCs requires the use of multiple markers for accurate identification. In this study, we developed a convolutional neural network-support vector machine (CNN-SVM) model, a supervised learning model that can classify CTCs by cell morphological features. To build the CNNSVM model, we employed the Wright-Giemsa staining method which stains the nucleus and membrane of cells. Through this method, the unique morphological characteristics of each cell were visualized, and it was mathematically brought into the CNN-SVM model. We selected two leukocyte cell lines (HL-60, Jurkat) and two breast cancer cell lines (MCF-7, MDA-MB-231) to apply the various cell composition in the blood for training of the model. Also, we prepared cancer cell aggregates that recapitulate CTC clusters using the double-spiral microfluidic chip we previously reported [1]. The CNN-SVM architecture consists of three layers: an input layer, a hidden layer, and an output layer displaying the visualized map, as shown in Figure 1(A). Figure 1(B) shows the structure of the hidden layer. Four morphological features were extracted and computed to determine the eigenvalue of each class. As a result, the true positive rate of cluster classes was over 90% which indicates developed CNN-SVM model accurately classifies the CTC cluster in various types of blood cells. Therefore, we believe this algorithm provides evidence for its potential as a powerful tool for cancer diagnosis and prognosis.