Cell Nuclei Segmentation Using Deep Learning Techniques

Pathological examination usually involves manual inspection of hematoxylin and eosin (H&E)- stained images, which is labor-intensive, prone to significant variations, and lacking reproducibility. One of the fundamental tasks to automate this process is to find all the cell nuclei in the H&E-stained images for further analysis. We attempt this problem using deep learning techniques. First, we introduce a semantic pixel-wise segmentation technique using dilated convolutions. We show that dilated convolutions are superior in extracting information from textured images. H&E-stained images are highly textured, which makes dilated convolutions an ideal technique to apply. Our dilated convolutional network (DCN) is constructed based on SegNet, a deep convolutional encoder-decoder architecture. Dilated convolution layers with increased dilation factors are used in the encoder to preserve image resolution. Dilated convolution layers with decreased dilation factors are used in the decoder to reduce gridding artifacts. Our DCN network was tested on synthetic data sets and a publicly available data set of H&E-stained images. We achieve better segmentation results than state-of-the-art. To further separate the instance of each cell nuclei, we adapt our DCN with a single shot multibox detector (SSD) and achieve promising results. Our methods are computationally efficient and can be run on a personal laptop computer. This work is the first step to wards using mathematical models to generate diagnostic inferences and providing clinically actionable knowledge to physicians and patients.