Generating Training Data for Ureter Segmentation Using Dual-Energy CT Two-Material Decomposition

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Urological Imaging

Background:

• Accurate ureter segmentation is crucial for diagnosing and managing various urological conditions.
• Deep learning models require large, high-quality datasets for effective training.
• Generating segmentation masks for ureters, especially on non-contrast CT, presents challenges.

Purpose of the Study:

• To assess the feasibility of using dual-energy CT (DECT)-based two-material decomposition to create training data for ureter segmentation.
• To develop and evaluate a deep learning model for ureter segmentation using DECT-derived virtual unenhanced (VUE) images.

Main Methods:

• A retrospective study involving 180 patients who underwent DECT urography.
• Virtual unenhanced (VUE) images were synthesized from late excretory phase (LEP) DECT images using two-material decomposition.
• Ground truth segmentation masks were generated on LEP images and paired with VUE images to form training datasets.
• A deep learning model (nnU-Net framework) was trained and validated on internal and external datasets.

Main Results:

• The internal test dataset achieved high performance: median Dice coefficient of 0.89, precision of 0.90, and recall of 0.88.
• External validation demonstrated limited performance: median Dice coefficient of 0.43 and recall of 0.28, with high precision (0.95).
• Statistically significant differences (P < 0.01) were observed in all metrics between internal and external validation datasets.

Conclusions:

• DECT-based two-material decomposition is a viable method for generating training data for ureter segmentation.
• The approach shows potential for ureter segmentation on non-contrast CT scans, despite limitations in external validation.
• Further research and multi-center validation are needed to improve generalizability and clinical applicability.