Urinary Stone Detection on CT Images Using Deep Convolutional Neural Networks: Evaluation of Model Performance and...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Urinary stone detection is crucial for patient management.
• Unenhanced CT is a standard imaging modality for urolithiasis.
• Developing automated detection tools can improve efficiency and accuracy.

Purpose of the Study:

• To assess the diagnostic accuracy of a cascading convolutional neural network (CNN) for detecting urinary stones on unenhanced CT scans.
• To evaluate the performance and generalizability of pretrained CNN models using labeled CT images across different scanners.

Main Methods:

• A retrospective study of 535 adult unenhanced abdominopelvic CT scans from two scanners.
• Development of a two-stage CNN: one for urinary tract segmentation, another for stone detection.
• Training and testing nine CNN model variations using different pretrained models (ImageNet, GrayNet, Random) and data sources (scanner 1, scanner 2, both).
• Evaluation using area under the receiver operating characteristic curve (AUC) and accuracy at the patient level.

Main Results:

• The GrayNet-pretrained model demonstrated superior performance compared to ImageNet-pretrained and Random-initialized models.
• Patient-level AUC for stone detection ranged from 0.92 to 0.95.
• The GrayNet model trained on both scanners (GrayNet-SB) achieved 95% accuracy, outperforming ImageNet-SB (91%) and Random-SB (88%).
• GrayNet-SB showed improved detection of smaller stones (<4 mm) and correctly identified all obstructive uropathy cases.

Conclusions:

• Cascading CNN models achieve high accuracy (AUC, 0.954) in detecting urinary tract stones on unenhanced CT.
• Transfer learning, utilizing datasets with labeled medical images, enhances CNN performance and cross-scanner generalizability.
• Automated detection systems show promise for improving the diagnosis of urolithiasis.