Automated Mouse Organ Segmentation Using Swin Transformers
Author Information
Author(s): Jiang Lu, Xu Di, Xu Qifan, Chatziioannou Arion, Iwamoto Keisuke S., Hui Susanta, Sheng Ke
Primary Institution: University of California San Francisco
Hypothesis
Can Swin UNEt TRansformers effectively segment mouse organs in micro-CT scans compared to existing methods?
Conclusion
Swin UNETR outperforms existing models in segmenting mouse organs, demonstrating robustness across different imaging conditions.
Supporting Evidence
- Swin UNETR consistently outperformed nnU-Net and AIMOS in terms of average dice similarity coefficient (DSC) and Hausdorff distance (HD95p).
- The model showed superior performance in segmenting organs like the bladder, lungs, and liver.
- Robustness was confirmed through testing on an independent dataset with different imaging protocols.
- Manual segmentation was often inaccurate, highlighting the need for automated methods.
Takeaway
This study shows that a new computer program can help quickly and accurately outline mouse organs in images, which is important for medical research.
Methodology
The study used public and in-house micro-CT datasets to train and evaluate the Swin UNETR model against nnU-Net and AIMOS.
Potential Biases
Inter-observer variations in manual annotations could introduce bias in model training and evaluation.
Limitations
The study's datasets were limited to 28 mice, and all scans were conducted in a prone position, which may affect generalizability.
Participant Demographics
The study involved genetically homogeneous mice, specifically 28 mice used across various imaging conditions.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website