Inspection of visible components in urine based on deep learning

Area of Science:

• Medical diagnostics
• Artificial intelligence in healthcare
• Urology and nephrology

Background:

• Urinary particle analysis is crucial for diagnosing nephropathy.
• Manual microscopy for urinary particles is time-consuming and subjective.
• Existing automated methods lack sufficient accuracy.

Purpose of the Study:

• To develop a deep learning-based method for accurate and efficient urinary particle analysis.
• To overcome the limitations of manual microscopy and current automated detection algorithms.

Main Methods:

• Utilized Resnet50 and Feature Pyramid Network (FPN) for feature extraction from urine microscopic images.
• Employed RetinaNet as the basic model for detecting seven cellular components.
• Investigated the impact of weight initialization and anchor scales on model performance.

Main Results:

• Achieved an accuracy of 88.65% for urinary particle detection.
• Processed images at a speed of 0.2 seconds per image using a GPU.
• Demonstrated performance comparable to two-stage algorithms with the speed of first-stage detectors.

Conclusions:

• A novel deep learning method for automated urinary particle analysis has been developed.
• This approach is expected to enhance automated urinalysis in clinical settings.
• The method shows potential for detecting other cell types in clinical diagnostics.