Study on Fine-Grained Visual Classification of Low-Resolution Urinary Erythrocyte.

Area of Science:

• Medical Diagnostics
• Biomedical Imaging
• Machine Learning in Healthcare

Background:

• Urine red blood cell morphology analysis is crucial for medical testing but current analyzers lack accuracy.
• Existing methods struggle with low image resolution, blurred features, and limited data, hindering fine-grained classification.
• This limits the widespread use of urine red blood cell morphology in medical examinations.

Purpose of the Study:

• To enhance the classification accuracy of low-resolution urine red blood cells.
• To address the limitations of current urine red blood cell morphology analyzers.
• To develop a practical reference for urine red blood cell morphology examination items.

Main Methods:

• Proposed a super-resolution method incorporating category-aware loss.
• Introduced an RBC-MIX data enhancement approach.
• Optimized cross-entropy loss for improved classification boundaries and feature distinction.

Main Results:

• Achieved a classification accuracy rate of 97.8% for low-resolution urine red blood cell images.
• Demonstrated outstanding classification performance using only category labels.
• The method effectively improves intra-class tightness and inter-class differences.

Conclusions:

• The developed super-resolution method significantly improves low-resolution urine red blood cell classification accuracy.
• This approach offers a practical and effective solution for urine red blood cell morphology examinations.
• The method shows potential for wider adoption in medical diagnostic testing.