Expert guidance and partially-labeled data collaboration for multi-organ segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Abdominal multi-organ segmentation in CT scans is vital for clinical applications.
• Current methods require extensive single-institution datasets or centralized multi-institution data, increasing labeling and collection burdens.
• Single-organ labels are more accessible and cost-effective than multi-organ labels.

Purpose of the Study:

• To develop an efficient collaborative framework for multi-organ segmentation using partially labeled data.
• To reduce the reliance on large, fully annotated datasets in medical image segmentation.
• To propose a method that leverages both single-organ and multi-organ labels effectively.

Main Methods:

• Proposed an expert-guided and partially-labeled data collaboration framework (EGPD-Seg).
• Introduced a reward-penalty loss function to focus on single-organ targets and mitigate unlabeled organ influence.
• Developed an expert-guided module for learning prior knowledge to segment unlabeled organs from single-organ labeled data.

Main Results:

• EGPD-Seg demonstrated effective multi-organ segmentation performance under partial label settings.
• The framework was validated on five diverse abdominal multi-organ segmentation datasets, including internal and external validation.
• The proposed modules interact to enhance segmentation accuracy with limited labels.

Conclusions:

• EGPD-Seg offers an effective solution for abdominal multi-organ segmentation with reduced annotation requirements.
• The collaborative mechanism between single and multi-organ labels, guided by expert input, significantly improves segmentation performance.
• This approach addresses the data challenges in medical image segmentation, making it more practical for clinical use.