Automatic identification of tumor cells for circulating tumor cells by convolutional neural networks

Liquid biopsy allows non-invasive collection of circulating tumor cells (CT -Cs) in blood without the need for sampling and can clearly demonstrate their presence in many types of cancer. In this study, we propose a method to automatically identify CTCs from fluorescence microscopy images and enable quantitative analysis based on convolu-tional neural networks (CNN). In this paper, a cell nucleus region cropping algorithm is applied in addition to a filtering process centered on a selective enhancement filter. Next, identification by SqueezeNet is performed. We performed the proposed method to 5,040 images of 6 samples and conducted experiments to identify CTCs. The number of detected CTCs was 148 (TPR = 100%), and the number of over-detected non-CTCs was 925. For the identification, TPR = 88.51% and FPR = 5.102% for the CNN model using SqueezeNet. The proposed method successfully reduced the number of detections by about 71.4% without missing any correct answers, but the proposed method did not show good results in any of the evaluation metrics.