Related Experiment Video
Updated: May 21, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
Deep learning-based segmentation of ultra-low-dose CT images using an optimized nnU-Net model
Yazdan Salimi1, Zahra Mansouri1, Chang Sun1,2
1Division of Nuclear Medicine and Molecular Imaging, Geneva University Hospital, CH-1211, Geneva, Switzerland.
La Radiologia Medica
|March 18, 2025
Summary
New deep learning models accurately segment organs on ultra-low-dose CT (LD-CT) scans, overcoming image quality limitations. These dedicated LD-CT models outperform traditional models, improving diagnostic accuracy in low-radiation imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Low-dose CT (LD-CT) protocols are crucial for various medical imaging applications, including emergency care and hybrid imaging.
- Image quality in LD-CT can be compromised, affecting the performance of standard deep learning (DL) segmentation models.
- Existing DL models are typically trained on high-quality images, with limited options for noisy LD-CT data.
Purpose of the Study:
- To develop a novel deep learning (DL) pipeline specifically for organ segmentation on ultra-low-dose CT (LD-CT) images.
- To address the challenge of reduced image quality in LD-CT, which impacts the accuracy of conventional segmentation models.
- To create dedicated DL models capable of robust organ segmentation even with significant noise and low radiation exposure.
Main Methods:
- Reconstruction of 274 CT datasets into full-dose (FD-CT) and simulated LD-CT images at varying radiation levels (1-10% of original).
- Training of new LD-nnU-Net models using LD-CT images, with existing FD-nnU-Net models on FD-CT images serving as references.
- Segmentation of bony tissues, soft tissues, and body contours using dedicated LD-nnU-Net models, followed by comparison with FD-CT reference masks and external LD-CT datasets.
Main Results:
- Standard FD-nnU-Net models showed decreased performance on LD-CT images, particularly at radiation doses below 10%.
- Newly developed LD-nnU-Net models achieved high average Dice scores: 0.937±0.049 for bony tissues, 0.905±0.117 for soft tissues, and 0.984±0.023 for body contour.
- The LD-nnU-Net models demonstrated superior segmentation performance compared to FD models when evaluated on external datasets of actual LD-CT images.
Conclusions:
- Conventional deep learning models trained on full-dose CT images perform inadequately on low-dose CT scans.
- Dedicated LD-nnU-Net models show significantly improved performance for organ segmentation on ultra-low-dose CT images.
- The developed LD-nnU-Net models enable accurate segmentation of noisy LD-CT images and are publicly available.
