Radiomics-Guided Deep Learning Networks Classify Differential Diagnosis of Parkinsonism.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Radiomics

Background:

• Differential diagnosis of atypical Parkinsonian syndromes is clinically challenging.
• ¹⁸F-fluorodeoxyglucose positron emission tomography (¹⁸F-FDG PET) is crucial for diagnosis.
• Existing methods may lack interpretability and precision.

Purpose of the Study:

• To develop and validate a radiomics-guided deep learning (DL) model for Parkinsonian syndrome diagnosis.
• To discover interpretable DL features for differential diagnosis.
• To enhance diagnostic accuracy and provide biological insights.

Main Methods:

• Recruited 1495 subjects (220 healthy controls, 1275 patients with IPD, MSA, or PSP) for ¹⁸F-FDG PET scans.
• Developed baseline radiomics and two DL models.
• Extracted DL latent features guided by radiomics for interpretability.

Main Results:

• The radiomics-guided DL model outperformed the baseline radiomics approach.
• DenseNet achieved high diagnostic sensitivity (95.7% IPD, 90.1% MSA, 91.2% PSP).
• DL latent features showed significant association with radiomics features and offered biological interpretations.

Conclusions:

• The radiomics-guided DL model provides interpretable insights for differentiating Parkinsonian disorders.
• This AI approach shows promise for personalized disease monitoring and diagnosis.
• Deep learning combined with radiomics enhances the diagnostic capabilities in neurodegenerative diseases.