Deep learning-based segmentation of the thorax in mouse micro-CT scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Small Animal Research

Background:

• Image-guided small animal irradiations require continuous anesthesia, impacting results and animal well-being.
• Automating organ contouring streamlines workflows, reducing anesthesia exposure and improving experimental reliability.

Purpose of the Study:

• To train and evaluate no-new-Net (nnU-Net) 2D and 3D U-Net architectures for automated thorax contouring in mouse micro-CT images.
• To assess the robustness of trained models on independent and external datasets, including contrast-enhanced CTs.
• To compare the performance of automated contouring against interobserver variability and manual contouring time.

Main Methods:

• Trained 2D and 3D U-Net architectures using the nnU-Net framework on native mouse micro-CT scans.
• Evaluated model performance on independent native CT datasets and external contrast-enhanced CT datasets.
• Assessed segmentation accuracy using mean surface distance (MSD) and 95th percentile Hausdorff distance (95p HD).
• Determined interobserver variability using the generalized conformity index and measured contouring time reduction.

Main Results:

• The 3D nnU-Net models demonstrated superior segmentation accuracy and robustness compared to 2D models.
• The best performing model (nnU-Net 3d_fullres) achieved MSD and 95p HD values below 1 mm (0.16 mm and 0.60 mm, respectively), meeting irradiation requirements.
• Contouring time was reduced by 98% with the 3D nnU-Net model compared to manual contouring.
• The 3D model's contouring accuracy favorably compared to interobserver variability.

Conclusions:

• Automated thorax contouring using 3D nnU-Net significantly enhances efficiency and accuracy in small animal image-guided irradiations.
• This automated approach minimizes anesthesia-related impacts, improving experimental outcomes and animal welfare.
• The 3D nnU-Net model provides a robust and accurate solution for contouring, outperforming previous methods and human variability.