Open-source domain adaptation to handle data shift for volumetric segmentation-use case kidney segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Current kidney segmentation models struggle with performance degradation due to variations in CT scan contrast phases.
• Differences in contrast uptake and kidney abnormalities present challenges for accurate kidney volume assessment.

Purpose of the Study:

• To develop a domain adaptation approach for robust kidney segmentation from CT volumes, independent of contrast dose and kidney function.
• To address domain shifts including contrast to non-contrast, arterial to venous phases, and normal to abnormal kidneys.

Main Methods:

• A domain adaptation technique using a latent space discriminator was employed.
• The model was trained on public non-contrast and arterial phase datasets.
• Validation was performed on diverse public and private datasets with varying contrast phases and kidney abnormalities.

Main Results:

• The proposed model achieved a 0.8892 DICE score across four datasets with domain shifts.
• It outperformed baseline models like TotalSegmentator and other domain adaptation methods on external validation.
• The approach demonstrated improved segmentation quality and required less data than baselines.

Conclusions:

• Domain adaptation significantly enhances kidney segmentation quality in CT scans.
• The developed model provides accurate kidney volume measurement irrespective of contrast variations or anomalies.
• An open-source codebase is available, facilitating clinical relevance and broader adoption for patient care.