Robust Automated Mouse Micro-CT Segmentation Using Swin UNEt TRansformers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pre-clinical Research

Background:

• Image-guided mouse irradiation is crucial for pre-clinical research before human trials.
• Accurate organ segmentation in micro-CT scans is vital for understanding radiation interventions.

Purpose of the Study:

• To segment mouse organs in micro-CT scans using Swin UNEt TRansformers (Swin UNETR).
• To benchmark Swin UNETR against 3D no-new-Net (nnU-Net) for segmentation accuracy and robustness.
• To evaluate the generalizability of Swin UNETR on external datasets with varying imaging conditions.

Main Methods:

• Swin UNETR was employed, treating mouse organ segmentation as a sequence-to-sequence task with a hierarchical Swin Transformer encoder and FCNN decoder.
• Models were trained and evaluated on open datasets, with further testing on an external dataset from a different micro-CT scanner.
• Performance was benchmarked using Dice Similarity Coefficient (DSC) and Hausdorff distance (HD95p).

Main Results:

• Swin UNETR demonstrated superior performance compared to nnU-Net and AIMOS in average DSC and HD95p.
• The model showed exceptional robustness and generalizability on the external dataset, handling variations in imaging noise and quality.
• Minor limitations were observed in intestine contouring for two specific mice.

Conclusions:

• Swin UNETR is a highly generalizable and efficient tool for automated organ contouring in pre-clinical workflows.
• The model's performance indicates its potential to significantly advance image-guided radiation studies.
• Swin UNETR offers a robust solution for segmentation tasks despite variations in micro-CT imaging parameters.