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 and methods. DICOM data (Digital Imaging and Communications in Medicine standard) from 41 patients with kidney neoplasms were used. Data included all phases of contrast-enhanced multispiral comp...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | Russian |
| Published: |
Federal State Autonomous Educational Institution of Higher Education I.M. Sechenov First Moscow State Medical University of the Ministry of Health of the Russian Federation (Sechenov University)
2023-03-01
|
| Series: | Сеченовский вестник |
| Subjects: | |
| Online Access: | https://www.sechenovmedj.com/jour/article/view/899 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1850050597075550208 |
|---|---|
| author | I. М. Chernenkiy M. M. Chernenkiy D. N. Fiev E. S. Sirota |
| author_facet | I. М. Chernenkiy M. M. Chernenkiy D. N. Fiev E. S. Sirota |
| author_sort | I. М. Chernenkiy |
| collection | DOAJ |
| description | Aim. Develop a neural network to build 3D models of kidney neoplasms and adjacent structures. Materials and methods. DICOM data (Digital Imaging and Communications in Medicine standard) from 41 patients with kidney neoplasms were used. Data included all phases of contrast-enhanced multispiral computed tomography. We split the data: 32 observations for the training set and 9 – for the validation set. At the labeling stage, the arterial, venous, and excretory phases were taken, affine registration was performed to jointly match the location of the kidneys, and noise was removed using a median filter and a non-local means filter. Then the masks of arteries, veins, ureters, kidney parenchyma and kidney neoplasms were marked. The model was the SegResNet architecture. To assess the quality of segmentation, the Dice score was compared with the AHNet, DynUNet models and with three variants of the nnU-Net (lowres, fullres, cascade) model. Results. On the validation subset, the values of the Dice score of the SegResNet architecture were: 0.89 for the normal parenchyma of the kidney, 0.58 for the kidney neoplasms, 0.86 for arteries, 0.80 for veins, 0.80 for ureters. The mean values of the Dice score for SegResNet, AHNet and DynUNet were 0.79; 0.67; and 0.75, respectively. When compared with the nnU-Net model, the Dice score was greater for the kidney parenchyma in SegResNet – 0.89 compared to three model variants: lowres – 0.69, fullres – 0.70, cascade – 0.69. At the same time, for the neoplasms of the parenchyma of the kidney, the Dice score was comparable: for SegResNet – 0.58, for nnU-Net fullres – 0.59; lowres and cascade had lower Dice score of 0.37 and 0.45, respectively. Conclusion. The resulting SegResNet neural network finds vessels and parenchyma well. Kidney neoplasms are more difficult to determine, possibly due to their small size and the presence of false alarms in the network. It is planned to increase the sample size to 300 observations and use post-processing operations to improve the model. |
| format | Article |
| id | doaj-art-b9b2b8dd6f894e79ab00c2b472da5d38 |
| institution | DOAJ |
| issn | 2218-7332 2658-3348 |
| language | Russian |
| publishDate | 2023-03-01 |
| publisher | Federal State Autonomous Educational Institution of Higher Education I.M. Sechenov First Moscow State Medical University of the Ministry of Health of the Russian Federation (Sechenov University) |
| record_format | Article |
| series | Сеченовский вестник |
| spelling | doaj-art-b9b2b8dd6f894e79ab00c2b472da5d382025-08-20T02:53:23ZrusFederal State Autonomous Educational Institution of Higher Education I.M. Sechenov First Moscow State Medical University of the Ministry of Health of the Russian Federation (Sechenov University)Сеченовский вестник2218-73322658-33482023-03-01141394910.47093/2218-7332.2023.14.1.39-49566Segmentation of renal structures based on contrast computed tomography scans using a convolutional neural networkI. М. Chernenkiy0M. M. Chernenkiy1D. N. Fiev2E. S. Sirota3Sechenov First Moscow State Medical University (Sechenov University)Sechenov First Moscow State Medical University (Sechenov University)Sechenov First Moscow State Medical University (Sechenov University)Sechenov First Moscow State Medical University (Sechenov University)Aim. Develop a neural network to build 3D models of kidney neoplasms and adjacent structures. Materials and methods. DICOM data (Digital Imaging and Communications in Medicine standard) from 41 patients with kidney neoplasms were used. Data included all phases of contrast-enhanced multispiral computed tomography. We split the data: 32 observations for the training set and 9 – for the validation set. At the labeling stage, the arterial, venous, and excretory phases were taken, affine registration was performed to jointly match the location of the kidneys, and noise was removed using a median filter and a non-local means filter. Then the masks of arteries, veins, ureters, kidney parenchyma and kidney neoplasms were marked. The model was the SegResNet architecture. To assess the quality of segmentation, the Dice score was compared with the AHNet, DynUNet models and with three variants of the nnU-Net (lowres, fullres, cascade) model. Results. On the validation subset, the values of the Dice score of the SegResNet architecture were: 0.89 for the normal parenchyma of the kidney, 0.58 for the kidney neoplasms, 0.86 for arteries, 0.80 for veins, 0.80 for ureters. The mean values of the Dice score for SegResNet, AHNet and DynUNet were 0.79; 0.67; and 0.75, respectively. When compared with the nnU-Net model, the Dice score was greater for the kidney parenchyma in SegResNet – 0.89 compared to three model variants: lowres – 0.69, fullres – 0.70, cascade – 0.69. At the same time, for the neoplasms of the parenchyma of the kidney, the Dice score was comparable: for SegResNet – 0.58, for nnU-Net fullres – 0.59; lowres and cascade had lower Dice score of 0.37 and 0.45, respectively. Conclusion. The resulting SegResNet neural network finds vessels and parenchyma well. Kidney neoplasms are more difficult to determine, possibly due to their small size and the presence of false alarms in the network. It is planned to increase the sample size to 300 observations and use post-processing operations to improve the model.https://www.sechenovmedj.com/jour/article/view/899neural networkkidney neoplasmsct scan3d modelingartificial intelligence |
| spellingShingle | I. М. Chernenkiy M. M. Chernenkiy D. N. Fiev E. S. Sirota Segmentation of renal structures based on contrast computed tomography scans using a convolutional neural network Сеченовский вестник neural network kidney neoplasms ct scan 3d modeling artificial intelligence |
| title | Segmentation of renal structures based on contrast computed tomography scans using a convolutional neural network |
| title_full | Segmentation of renal structures based on contrast computed tomography scans using a convolutional neural network |
| title_fullStr | Segmentation of renal structures based on contrast computed tomography scans using a convolutional neural network |
| title_full_unstemmed | Segmentation of renal structures based on contrast computed tomography scans using a convolutional neural network |
| title_short | Segmentation of renal structures based on contrast computed tomography scans using a convolutional neural network |
| title_sort | segmentation of renal structures based on contrast computed tomography scans using a convolutional neural network |
| topic | neural network kidney neoplasms ct scan 3d modeling artificial intelligence |
| url | https://www.sechenovmedj.com/jour/article/view/899 |
| work_keys_str_mv | AT imchernenkiy segmentationofrenalstructuresbasedoncontrastcomputedtomographyscansusingaconvolutionalneuralnetwork AT mmchernenkiy segmentationofrenalstructuresbasedoncontrastcomputedtomographyscansusingaconvolutionalneuralnetwork AT dnfiev segmentationofrenalstructuresbasedoncontrastcomputedtomographyscansusingaconvolutionalneuralnetwork AT essirota segmentationofrenalstructuresbasedoncontrastcomputedtomographyscansusingaconvolutionalneuralnetwork |