Automated Detection of Circulating Tumor Cells Using Faster Region Convolution Neural Network

Reliable detection and numeration of circulating tumor cells (CTCs) holds great promise for personalizing medicine and treatment effectiveness monitoring. However, traditional manual CTCs detection methods rely on morphological knowledge of cytologists are time consuming, and results are low in specificity, and sensitivity. In this work, an automated and robust CTCs detection method is proposed based on scanned microscopy images of peripheral blood samples. Firstly, a CTCs database is constructed from optical microscopy images collected from a local hospital and manually annotated by cytologists. Then, we propose to fine tuning a pre-trained Faster R-CNN network in end-to-end manner for high-level feature extraction and automatic detection of CTCs, which demonstrates high computational efficacy. Finally, extensive experiments are conducted on the CTCs database to evaluate the effectiveness of the proposed method, where an averaged F1 score of 0.855 is reported. Quantitative comparison with the state of the art method reveals the competitiveness of the proposed Faster R-CNN detector in real clinical practice.