Convolutional neural network-based kidney volume estimation from low-dose unenhanced computed tomography scans.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Renal Disease Management

Background:

• Accurate kidney volume measurement is crucial for diagnosing and managing renal diseases.
• Current semi-automated methods for kidney volume determination are time-consuming and error-prone.
• Automated methods often require contrast-enhanced CT scans, limiting their applicability.

Purpose of the Study:

• To develop an automated method for estimating kidney volume using non-contrast, low-dose CT scans.
• To assess the feasibility and accuracy of Convolutional Neural Network (CNN) models for kidney segmentation in urolithiasis patients.
• To provide a faster and more reproducible alternative to current kidney volume estimation techniques.

Main Methods:

• 2D CNN models were trained on manually segmented kidney images from low-dose, unenhanced CT scans of 210 patients.
• Segmentation accuracy was evaluated using the Dice Similarity Coefficient (DSC).
• The automated method was validated against semi-automated measurements from radiologists on 22 unseen cases.

Main Results:

• The CNN-enabled kidney volume estimation averaged 32 seconds per scan for both kidneys.
• High segmentation accuracy was achieved with DSC values of 0.91 for the left kidney and 0.86 for the right kidney.
• Excellent agreement was found between automated and semi-automated volume estimations (ICC > 0.89).

Conclusions:

• CNN-enabled kidney volume estimation is a feasible and highly reproducible technique.
• This automated approach is effective even with low-dose, non-enhanced CT scans.
• Automatic kidney segmentation can significantly enhance the quantitative value of radiological reports.