Label-free classification of circulating tumor cell clusters using convolutional neural network for image analysis

Abstract

Metastatic cancer has been known to be induced by single circulating tumor cells (CTCs) or CTC clusters derived from a primary tumor, and as a major cause of death worldwide. Recent studies have revealed that CTC clusters play a key role in cancer metastasis rather than single CTCs [1]. Once the CTC clusters are isolated from the patient's blood, the immunofluorescence staining methods are generally used to classify and analyze CTC clusters expressing specific biomarkers. However, the staining process has limitations of being timeconsuming and labor-intensive. In addition, essential washing steps within the procedure cause cell loss, and given the rarity of CTC clusters (0 to 5 in 10 ml of whole blood), this may limit the validity of CTC cluster assays. Here, to overcome the above limitations we developed a deep learning model that can rapidly classify CTC clusters and define cluster types with a label-free method. The validation accuracy of CTC cluster classification was calculated and compared in AlexNet Convolution Neural Network (CNN) model, SqueezeNet CNN model, and hyperparameter pooling layer added SVM (Support Vector Machine) model, respectively. Among the three models, the highest validation accuracy rate of 98% was obtained when using the SVM model. In addition, homotypic and heterotypic CTC clusters were successfully discriminated against. This learning model can be used for predicting the prognosis of cancer patients through morphological information of CTC clusters and classification of cluster types.