Automatic Detection of Circulating Tumor Cells with Very Deep Residual Networks

Abstract

Automatic detection of circulating tumor cells (CTCs) in microscopic images is a very challenging task due to the variable artificial and environmental factors, such as inconsistency of light intensity and staining, cell adhesion, multiple impurities and so on. In order to meet these challenges, we propose a novel deep multiscale residual network (DMRN) for CTCs detection. Compared with existing methods either low-level hand-crafted features or CNNs with shallower architectures, our deep networks can acquire more discriminative features for more accurate detection. To train very deep networks more efficiently, we propose a set of schemes to ensure effective training and learning under limited training data. First, we apply the residual learning to generate more discriminative features and overcome the overfitting problem when a network goes to deeper. Then, a fully residual convolutional network (FRCN) is proposed to produce the prediction maps of CTCs. Finally, we propose to integrate multi-scale contextual information in proposed FRCN and fuse these prediction maps both global and local features of CTCs, making the prediction more accurate and robust. We built three DMRN models to study the impact of network depth on model performance. Each model was tested on our own dataset containing complex jamming information. The DRMN-50 model which has a depth of 50 layers performs best among three models with Jaccard-index of 0.810 (with a pixel accuracy of 99.8% as a reference index) and its performance outperform other existing state-of-art methods such as U-Net in other domain. The result also depicts the accurate and robust performance of proposed method in complex environment.