Comparative analysis of deep learning neural networks for the segmentation of cancer cell nuclei on immunohistochemical fluorescent images

The analysis of histological and immunohistochemical images forms the basis for the diagnosis of many types of cancer. The process of automating the analysis of digital images, in particular, the segmentation of cell nuclei on them, is of great attention recently. Due to the excellent performance of deep learning neural networks and the relatively high level of reliability of the obtained results, it becomes possible to combine manual and automated image processing. To date, many neural network architectures have been created for segmenting objects in images. However, the high variability of images of cancer cells does not allow creating an universal algorithm for segmenting the cells nuclei on images of different types of tissues obtained using different techniques. In this paper, a comparative analysis of the architectures of deep learning neural networks for segmentation of cancer cell nuclei on immunohistochemical fluorescent images of breast cancer was carried out. It was established that networks based on the U-Net architecture give consistently good results. The UNet 3+ architecture showed the best segmentation quality.