Exploring Automated Contouring Across Institutional Boundaries: A Deep Learning Approach with Mouse Micro-CT Datasets

Area of Science:

• Medical imaging
• Artificial intelligence in preclinical research
• Radiotherapy research

Background:

• Accurate organ segmentation in mouse models is crucial for preclinical radiation studies.
• Current segmentation methods may lack robustness across different imaging conditions.

Purpose of the Study:

• To evaluate Swin UNEt Transformers (Swin UNETR) for segmenting mouse organs in micro-computed tomography (micro-CT) scans.
• To benchmark Swin UNETR against 3D no-new-Net (nnU-Net) and other models.
• To assess the robustness and generalizability of Swin UNETR on external datasets.

Main Methods:

• Swin UNETR was employed for sequence-to-sequence organ segmentation.
• A hierarchical Swin Transformer encoder and a Fully Convolutional Neural Network (FCNN) decoder with skip connections were utilized.
• Models were trained and validated on open datasets, with further testing on an external dataset with varying imaging parameters.

Main Results:

• Swin UNETR demonstrated superior performance over nnU-Net and AIMOS in average Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD95p).
• The model showed exceptional robustness and generalizability on an external dataset with lower image quality and higher noise.
• Minor limitations were observed in intestine contouring for specific cases.

Conclusions:

• Swin UNETR is a highly generalizable and efficient tool for automated organ contouring in pre-clinical workflows.
• The model's performance indicates its suitability for image-guided mouse irradiation studies.
• Swin UNETR offers improved accuracy and reliability compared to existing methods, especially under challenging imaging conditions.