DECTGoutSys: Reducing False Positive Gout Diagnoses via a Machine Vision Pipeline for Crystal Tophi...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Rheumatology

Background:

• Gout is a leading chronic inflammatory arthritis globally.
• Dual-energy computed tomography (DECT) visualizes monosodium urate (MSU) crystal deposits in gout with high sensitivity.
• DECT may produce artifacts in healthy individuals, reducing diagnostic specificity.

Purpose of the Study:

• To develop and validate an automated system for localizing and classifying MSU crystal deposits from DECT scans.
• To differentiate between gouty tophi and imaging artifacts.
• To improve the specificity of gout diagnosis using machine learning.

Main Methods:

• A three-stage automated system combining computer vision and deep learning models.
• Image processing to crop and filter regions of interest (RoIs) from DECT scans.
• Three deep learning models fine-tuned to classify RoIs based on size (small, medium, large).
• A machine learning system to classify patients as gout or control based on RoI classifier outcomes.

Main Results:

• The pipeline achieved 91.89% patient-level diagnostic accuracy, 87.23% sensitivity, and 100.00% specificity.
• The best ROC AUC at the patient level was 97.16%.
• RoI-level classifiers demonstrated 89.61% accuracy, 85.42% sensitivity, and 93.70% specificity.

Conclusions:

• Machine and deep learning can significantly enhance the specificity of gout diagnosis from DECT scans.
• The automated system effectively distinguishes gouty tophi from artifacts, maintaining good sensitivity.
• This AI-driven approach offers a promising tool for accurate and specific gout diagnosis.