Kidney, ureter, and urinary bladder segmentation based on non-contrast enhanced computed tomography images using...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation of the urinary system is crucial for diagnosing diseases on non-contrast computed tomography (CT).
• Existing segmentation methods may require improvement for clinical application in urinary tract disease diagnosis.

Purpose of the Study:

• To develop and evaluate a modified U-Net model for segmenting the urinary system (kidneys, ureters, urinary bladder) on non-contrast CT scans.
• To establish a foundation for machine learning-based detection of urinary system lesions.

Main Methods:

• Utilized non-contrast abdominopelvic CT scans from patients diagnosed with urinary stones.
• Applied region of interest extraction followed by urinary system segmentation using a modified U-Net architecture.
• Evaluated model robustness and performance using fivefold cross-validation and a separate test dataset.

Main Results:

• Achieved an average Dice coefficient of 0.8673 in fivefold cross-validation for whole urinary system segmentation.
• Specific Dice coefficients were 0.9651 for kidneys, 0.7172 for ureters, and 0.9196 for the urinary bladder.
• The best-performing model on the test dataset yielded an average Dice coefficient of 0.8623 for the whole system.

Conclusions:

• The modified U-Net model effectively segments the urinary system in non-contrast CT.
• This segmentation technique provides a basis for machine learning-based detection of urinary tract pathologies.
• The model shows promise for improving the diagnosis of conditions like kidney stones and bladder tumors.