Faster R-CNN based Robust Circulating Tumor Cells Detection with Improved Sensitivity
Yunxia Liu, Hongchao Si, Yingjie Chen, Yuehui Chen · 2019
Automatic detection and numeration of circulating tumor cells (CTCs) in scanned microscopic images of peripheral blood provides effective auxiliary means for clinical diagnosis of cancer, individualized treatment, prognosis judgement evaluation and so on. There have been some work aims at reducing the subjectivity and labor intensity of cytologists with machine learning methods. In this paper, a robust CTCs detection algorithm based on improved Faster R-CNN network is proposed. Special efforts have been devoted to detection of single CTC for improved sensitivity, which is more favorable in real applications. Firstly, a whole slide image preprocessing algorithm is proposed based on statistical analysis of the self-constructed CTCs database. Then, an anchor adjustment and extension method is proposed to alleviate the imbalance between limited number of positive CTCs examples and huge amount of negative normal cells. Experimental results carried out on the self-established CTCs database show that the proposed algorithm demonstrates improved performance in CTCs detection.