Classification of circulating tumor cell clusters by morphological characteristics using convolutional neural network-support vector machine

Metastasis is the leading cause of cancer-associated deaths, and the circulating tumor cell (CTC) cluster plays a significant role as a precursor to metastasis. Thus, there is a great demand for high-throughput identification of rare CTC clusters for prognostic diagnosis. Immunofluorescence staining is considered the gold standard for identifying CTCs. However, as CTC clusters are extremely heterogeneous cells, multiple staining markers are required for accurate discrimination. Additionally, the staining procedure is tedious and the analysis of large amounts of stained images is labor-intensive and error-prone. Recently, machine learning-based identification has been introduced to achieve accurate discrimination, but they still rely on immunofluorescence staining for dataset preparation. In this study, we developed a hybrid algorithm, a convolutional neural network support vector machine (CNN-SVM), for the accurate classification of CTC clusters without immunofluorescence staining. In dataset preparation, the Wright-Giemsa staining was used to highlight the morphological features of the cells. Four morphological characteristics that display the unique traits of cells were drawn with each eigenvector, as a result of learning, the algorithm classified CTC clusters of various configurations with a sensitivity and specificity of > 90%. Therefore, our algorithm is expected to be a powerful tool for cancer diagnosis and prognosis.

Automated summary

In this study, a hybrid algorithm was developed for accurate classification of CTC clusters without immunofluorescence staining, providing a powerful tool for cancer diagnosis and prognosis.