Towards Generic Abdominal Multi-Organ Segmentation with multiple partially labeled datasets.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Publicly available medical datasets enable universal segmentation model development.
• Current methods struggle with partially labeled data, cross-site domain shifts, and underutilization of unlabeled data.

Purpose of the Study:

• To introduce a novel framework, GAMOS (Generic Abdominal Multi-Organ Segmentation), for robust medical image segmentation.
• To address challenges in partial labeling, domain shifts, and leveraging unlabeled data in medical segmentation tasks.

Main Methods:

• GAMOS integrates diffusion models with a self-guidance strategy for partial labeling.
• A self-distillation mechanism is employed to utilize unlabeled data effectively.
• Sparse semantic memory and a sparse similarity loss mitigate domain shifts and enhance representation consistency.

Main Results:

• GAMOS achieved a mean Dice Similarity Coefficient (DSC) of 91.33% and a mean 95th percentile Hausdorff Distance (HD95) of 1.83 on labeled foreground regions.
• For unlabeled foreground regions, GAMOS obtained a mean DSC of 86.88% and a mean HD95 of 3.85.
• The framework demonstrated superior performance and generalization ability compared to state-of-the-art methods.

Conclusions:

• GAMOS offers a powerful solution for generic abdominal multi-organ segmentation, excelling with partially labeled and unlabeled data.
• The proposed methods effectively address domain shifts and improve the utilization of diverse medical imaging datasets.