Segmentation of renal structures based on contrast computed tomography scans using a convolutional neural network

Aim. Develop a neural network to build 3D models of kidney neoplasms and adjacent structures. Materials methods. DICOM data (Digital Imaging Communications in Medicine standard) from 41 patients with were used. Data included all phases contrast-enhanced multispiral computed tomography. We split the data: 32 observations for training set 9 – validation set. At labeling stage, arterial, venous, excretory taken, affine registration was performed jointly match location kidneys, noise removed using median filter non-local means filter. Then masks arteries, veins, ureters, parenchyma marked. The model SegResNet architecture. To assess quality segmentation, Dice score compared AHNet, DynUNet three variants nnU-Net (lowres, fullres, cascade) model. Results. On subset, values architecture were: 0.89 normal kidney, 0.58 neoplasms, 0.86 0.80 ureters. mean SegResNet, AHNet 0.79; 0.67; 0.75, respectively. When model, greater variants: lowres 0.69, fullres 0.70, cascade 0.69. same time, comparable: 0.58, 0.59; had lower 0.37 0.45, Conclusion. resulting finds vessels well. Kidney are more difficult determine, possibly due their small size presence false alarms network. It is planned increase sample 300 use post-processing operations improve