A Machine Learning Model Based on Radiomic Features as a Tool to Identify Active Giant Cell Arteritis on [¹⁸F]FDG-PET...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Giant cell arteritis (GCA) diagnosis and therapy monitoring rely on [18F]FDG-PET/CT scans.
• Differentiating active GCA from atherosclerosis in follow-up scans can be challenging.
• Machine learning (ML) offers potential for objective image analysis.

Purpose of the Study:

• To assess the feasibility of an ML model using radiomic features for active GCA detection in aorta on [18F]FDG-PET/CT.
• To differentiate active GCA from atherosclerosis in follow-up scans for therapy monitoring.
• To compare ML model performance against clinical reports.

Main Methods:

• Retrospective analysis of 64 [18F]FDG-PET scans from GCA patients and controls.
• Extraction of 95 radiomic features from delineated aortic segments.
• Training and validation of 441 ML models using various feature selection and classification methods.
• Performance evaluation using Area Under the Curve (AUC) and comparison with clinical reports.

Main Results:

• The optimal ML model (10 features, ANOVA, random forest) achieved an AUC of 0.92 ± 0.01.
• The ML model demonstrated superior performance over clinical reports in detecting active GCA (PPV 0.83 vs. 0.80, NPV 0.85 vs. 0.79, accuracy 0.84 vs. 0.79).
• Explainability analysis using occlusion maps identified key aortic regions for ML decision-making.

Conclusions:

• A radiomics-based ML model can accurately identify active GCA in the aorta using [18F]FDG-PET/CT scans.
• The ML model effectively differentiates GCA from atherosclerosis, showing potential as a monitoring tool.
• This approach may enhance the interpretation of challenging [18F]FDG-PET/CT scans in GCA patients.