Biomarkers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Deep learning for Parkinson's disease (PD) diagnosis via MRI is hindered by limited labeled datasets.
• Transfer learning offers a solution by leveraging knowledge from pre-trained models on large medical imaging datasets.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning using a pre-trained 3D convolutional neural network (Med3D) for PD detection from MRI scans.
• To address challenges of limited data and class imbalance in PD diagnosis.

Main Methods:

• Utilized Med3D, a 3D CNN pre-trained on diverse medical segmentation tasks, for feature extraction from 3D brain MRI scans.
• Adapted Med3D for PD classification, fine-tuned the network end-to-end, and employed sampling techniques for class imbalance.
• Trained and evaluated the model on the Parkinson's Progression Markers Initiative (PPMI) database.

Main Results:

• Achieved 94.03% accuracy, 100% sensitivity, 60% specificity, 93.44% precision, and 0.966 F1-score on the test set (AUC-ROC of 0.8).
• Significantly outperformed a baseline model trained from scratch, which achieved 75% accuracy.
• Demonstrated the efficacy of pre-trained weights for enhancing PD detection performance.

Conclusions:

• Transfer learning with Med3D provides a highly effective framework for PD detection using MRI.
• The approach shows significant improvement over traditional methods, highlighting the value of pre-trained models.
• Establishes a promising foundation for developing reliable computer-aided diagnosis systems for PD and other neurodegenerative diseases.