Classification of Parkinsonian syndromes from FDG-PET brain data using decision trees with SSM/PCA features

Area of Science:

• Neurology
• Medical Imaging
• Machine Learning

Background:

• Neurodegenerative diseases like Parkinsonian syndromes pose diagnostic challenges.
• Medical imaging, specifically fluorodeoxyglucose positron emission tomography (FDG-PET), is crucial for differential diagnosis.
• Accurate classification aids in timely and appropriate patient management.

Purpose of the Study:

• To classify FDG-PET brain scans for differentiating Parkinsonian syndromes (Parkinson's disease, multiple system atrophy, progressive supranuclear palsy) from healthy controls.
• To evaluate the effectiveness of the scaled subprofile model/principal component analysis (SSM/PCA) combined with C4.5 decision tree classification for this purpose.

Main Methods:

• FDG-PET brain image data from patients and controls were analyzed using SSM/PCA to extract key features.
• Subject scores derived from SSM/PCA were used as input for a supervised C4.5 decision tree classifier.
• Leave-one-out cross-validation was employed to assess classifier performance, with comparisons to other classification methods.

Main Results:

• The C4.5 decision tree classifier demonstrated effective classification of FDG-PET scans for differentiating Parkinsonian syndromes.
• Decision tree classification offers easily interpretable results, crucial for clinical medical diagnosis.
• Visualizing decision trees aids in understanding the classification process.

Conclusions:

• FDG-PET imaging, analyzed with SSM/PCA and C4.5 decision trees, is a viable method for the differential diagnosis of Parkinsonian syndromes.
• The interpretability of decision trees is a significant advantage in clinical settings.
• Future work could enhance accuracy by increasing training data, employing ensemble methods like bagging, and integrating functional MRI (fMRI) data.