An Algorithm for Segmentation of Kidney Tissues on CT Images Based on a U-Net Convolutional Neural Network

The effects of different image preprocessing operations, as well as various parameters of the U-Net architecture neural network, on the accuracy of segmentation of kidney tissues in CT images were studied. The greatest segmentation accuracy was found to be obtained by sequential application to each image of two-level brightness windowing, specification of the brightness histogram, and bringing the dynamic range of brightness to the interval [–1, 1]. In contrast to existing approaches to image segmentation using neural networks, two-level windowing in the present study was followed by assignment of random values in the range from –1 to 1 with a uniform probability distribution to background pixels. In addition, the operation of masking areas of the image was used, which provides for increases in the stability of neural network training.