Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Renal structures CECT segmentation

Link

https://doi.org/10.47093/2218-7332.2023.14.1.39-49

Indexing

Keywords: Neural Network, Kidney Neoplasms, Computed Tomography, 3D Modeling, Image Segmentation, Renal Arteries, Renal Veins, Ureters, Kidney Parenchyma

Content: CT, GU, OI, BQ

Author(s)

Ivan M. Chernenkiy

Michail M. Chernenkiy

Dmitry N. Fiev

Evgeny S. Sirota

Organization(s)

Sechenov First Moscow State Medical University (Sechenov University)

Contact

chernenkiy_i_m@stafF.sechenov.ru

Funding

The study was not sponsored (own resources).

Ethical review

The study was approved by the Local Ethics Committee of Sechenov University.

Comments

The study was not sponsored (own resources). Authors express their gratitude to the Institute of Computer Science and Mathematical Modeling of Sechenov University for access to the computing cluster.

Date

Published: 2023-03-30

Model

Architecture

SegResNet architecture, a type of UNet convolutional neural network.

Availability

https://github.com/blacky-i/nephro-segmentation

Clinical benefit

Enables 3D surgical planning, preoperative planning, intraoperative navigation, and determination of optimal surgical access for kidney neoplasms.

Clinical workflow phase

Pre-operative planning

Degree of automation

Automated 3D model construction.

Indications for use

Segmentation of kidney neoplasms and adjacent renal structures for 3D modeling in surgical planning.

Input

DICOM data from contrast-enhanced multispiral computed tomography, including arterial, venous, and excretory phases.

Instructions

Pre-processing included affine registration, median filter, and non-local means filter. Augmentation methods used were Zoom (0.9-1.1 scale, 0.3 probability), Rotate 90 (0.3 probability), Flip (0.3 probability per axis), Adjust Contrast (0.5-4.5 scale, 0.3 probability), and Histogram Shift (0.3 probability). Dice coefficient was used as the loss function. Novograd optimizer was used with a batch size of 4 and an input window of [96, 96, 96] voxels.

Limitations

Difficulty in determining kidney neoplasms due to small size and false positives. The model may identify formations outside the organ. Limited by a small dataset size (41 observations) for training.

Output

Description: Segmentation masks for arteries, veins, kidney parenchyma, kidney neoplasms, ureters, and background. 3D models of kidney neoplasms and adjacent structures.

Recommendation

Increase sample size to 300 observations. Implement post-processing operations to improve model quality.

Reproducibility

Code for reproducing results is available on GitHub.

Use

Intended: Image segmentation, 3D modeling, Surgical planning, Intraoperative navigation

User

Intended: Urologist, Radiologist, Surgeon, Researcher, IT specialist