Automatic Detection of Tumor Cells in Microscopic Images of Unstained Blood using Convolutional Neural Networks

Accessible high-performance computing power has recently spiked interest in medical image analysis and processing. Biomedical image segmentation has been used to aid in the process of medical analysis and diagnosis. In this paper we present a novel approach to identifying circulating tumor cells (CTCs) using convolutional neural networks on Dark Field microscopic images of unstained blood. We use a modified U-Net that is able to automatically perform image segmentation in order to detect CTCs. We perform detection on our own dataset containing input images and ground truth label images. Detection is done on small image patches using a sliding window mechanism in order to reduce computation time. The final result is reconstructed from the patches and further refined using post-processing. The total number of CTCs is computed from the segmented image using the Hough circle algorithm. We were able to obtain over 99.8% accuracy using our data set.