.png){kind=icon}
.png){kind=icon}
Return
Organ Segmentation in CT Images Using nnU-Net and 3D Model Generation With Marching Cubes
.png){kind=icon}
PRE
AI
Y
F
A
M
DOI：10.1002/ima.70413.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multiorgan segmentation from CT scans is crucial for numerous clinical and modeling applications, yet manual delineation remains labor-intensive. This study introduces a fully automated pipeline that integrates self-configuring nnU-Net-based multiorgan segmentation with automated 3D mesh generation and quantitative mesh quality assessment on the official CT-ORG benchmark. A 3D full-resolution nnU-Net v2 model was developed using 119 contrast-enhanced and noncontrast-enhanced CT scans from the CT-ORG dataset. The workflow comprised automated preprocessing, data augmentation, and fivefold cross-validation, followed by organ-surface reconstruction using Marching Cubes and Laplacian smoothing. Segmentation performance was evaluated on the predefined independent test set of 21 CT scans, and the generated surface models were assessed using quantitative mesh quality metrics. Fold 4 achieved the highest validation Dice score of 0.951 and was selected for final inference. Dice and IoU scores on the independent test set were as follows: liver (0.949, 0.904), lungs (0.975, 0.952), kidneys (0.911, 0.841), skeletal structures (0.899, 0.818), and urinary bladder (0.865, 0.778). Validation curves demonstrated stable learning and minimal overfitting. The developed framework provides a reliable and efficient solution for automated CT-based multiorgan segmentation and mesh generation. It reduces clinical workload and accelerates the production of 3D anatomical models, offering utility in surgical planning, medical education, and patient-specific phantom creation.
Keywords:
3D medical modelling
computed tomography (CT)
deep learning
multiorgan segmentation
Nnu-net
Journal
IF:
2.5
Papers:
2.1K
Citations:
2.3K
Organization
Cited Papers