An End-to-End System for Automatic Urinary Particle Recognition with Convolutional Neural Network.

Area of Science:

• Medical diagnostics
• Computer vision
• Machine learning

Background:

• Manual urine sediment analysis is crucial for diagnosing renal and urinary tract diseases but is time-consuming and subjective.
• Traditional automated methods rely on hand-crafted features, limiting recognition accuracy.
• Automated analysis of microscopic urine particles is needed to improve diagnostic efficiency.

Purpose of the Study:

• To develop and evaluate deep learning models for automated recognition of urinary particles.
• To compare the performance of CNN-based object detection methods for this task.
• To optimize CNN approaches for accurate and efficient urine sediment analysis.

Main Methods:

• Utilized convolutional neural networks (CNNs) for end-to-end feature learning, avoiding hand-crafted features.
• Applied state-of-the-art object detection models, Faster R-CNN and Single Shot Multibox Detector (SSD), and their variants.
• Trained and evaluated models on a dataset of 5,376 annotated microscopic urine images across 7 particle categories.

Main Results:

• Achieved a best mean average precision (mAP) of 84.1% for urinary particle recognition.
• Demonstrated rapid processing, with an average inference time of 72 ms per image on a NVIDIA Titan X GPU.
• CNN-based methods significantly outperformed traditional approaches in accuracy and efficiency.

Conclusions:

• Convolutional neural networks offer a powerful and efficient solution for automated urine sediment analysis.
• Deep learning-based object detection can accurately identify various urinary particles, aiding clinical diagnosis.
• This approach enhances the speed and objectivity of urine particle analysis, supporting better patient evaluation.