Automated CT segmentation for lower extremity tissues in lymphedema evaluation using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Clinical assessment of lymphedema, especially its severity and fibrotic changes, is difficult with current methods.
• Accurate tissue segmentation in lower extremity CT scans is crucial for quantitative lymphedema evaluation.

Purpose of the Study:

• To develop and validate a deep learning (DL) segmentation tool for automated analysis of tissue components in lower extremity CT scans.
• To enable precise measurement of lymphedema severity and fibrotic lesions.

Main Methods:

• A DL model (Unet++ with EfficientNet-B7 encoder) was trained on CT venography scans from 118 patients with gynecologic cancers.
• Segmentation of fat, muscle, and fluid-fibrotic tissue was performed using 3D slicer as a reference standard.
• The DL model's accuracy was validated using Dice Similarity Coefficient (DSC) and Volumetric Similarity (VS) in internal and external datasets.
• A graphical user interface (GUI) was developed for visualizing segmentation results.

Main Results:

• The DL algorithm demonstrated high segmentation accuracy, with mean DSCs ranging from 0.945 to 0.999 and mean VSs from 0.97 to 0.999 across validation sets.
• Volumetric analysis showed no significant differences between DL measurements and reference standards for total leg volume or component volumes (p > 0.05).
• The developed GUI effectively maps lymphedema by highlighting segmented tissue components.

Conclusions:

• The DL algorithm provides an automated and accurate tool for segmenting and quantifying lower extremity tissue components on CT scans.
• This tool facilitates automated lymphedema evaluation, severity assessment, and mapping.
• The high accuracy of the DL tool offers significant clinical relevance for managing lymphedema.