Clinical-Inspired Framework for Automatic Kidney Stone Recognition and Analysis on Transverse CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Urology

Background:

• Accurate kidney stone diagnosis from CT images is crucial for patient care.
• Existing automated methods often neglect valuable clinical knowledge, limiting performance and interpretability.

Purpose of the Study:

• To develop a clinical-inspired framework for kidney stone detection and evaluation in CT images.
• To integrate the diagnostic process of urologists into an automated system.

Main Methods:

• A three-module framework was designed: object attention for localization, feature-driven discrimination for identification, and a clustering/graphic combination analysis for evaluation.
• The framework incorporates clinical decision-making steps of urologists.
• A custom dataset of 27,885 CT images was utilized.

Main Results:

• The object attention module showed a 1% improvement over Yolov7 for localization.
• The analysis module achieved 21.9% higher average cluster accuracy and 17.35% lower average error compared to AR-DBSCAN and formula methods.
• The framework demonstrated superior performance in kidney stone recognition and evaluation.

Conclusions:

• The proposed clinical-inspired framework significantly enhances kidney stone recognition and evaluation in CT images.
• Mimicking the urologist's diagnostic process leads to a more effective and interpretable automated solution.
• This approach represents a state-of-the-art solution for kidney stone analysis.