Subcutaneous edema segmentation on abdominal CT using multi-class labels and iterative annotation.

Area of Science:

• Medical imaging analysis
• Computational pathology
• Artificial intelligence in medicine

Background:

• Anasarca, characterized by widespread edema due to organ dysfunction, requires accurate quantification for clinical management.
• Non-invasive estimation of edema volume using abdominal CT scans holds clinical potential.
• Edema segmentation is challenging due to complex visual characteristics and limited annotated data.

Purpose of the Study:

• To develop an accurate, non-invasive method for estimating edema volume from abdominal CT scans.
• To minimize false positives in edema segmentation.
• To address the challenges posed by complex edema appearance and lack of annotated volumes.

Main Methods:

• A weakly supervised learning approach for edema segmentation was proposed.
• Utilized initial edema labels from the Intensity Prior method and surrounding tissue labels as anatomical priors.
• Employed a multi-class 3D nnU-Net segmentation network with an iterative annotation workflow.

Main Results:

• The proposed method achieved comparable segmentation accuracy to Intensity Prior (Dice Similarity Coefficient: 61.5% vs. 61.7%).
• Significantly reduced the False Positive Rate from 1.8% to 1.1% (p < 0.001).
• Automated edema volumes strongly correlated with manual annotations (R² = 0.87).

Conclusions:

• Weakly supervised learning with 3D multi-class labels and iterative annotation enables high-quality edema segmentation with minimal false positives.
• Automated edema segmentation provides reliable edema volume estimates.
• The approach shows promise for clinical monitoring of anasarca.