Deep Learning-Based Brainstem Segmentation and Multi-Class Classification for Parkinsonian Syndrome.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Structural MRI is key for identifying atrophy in Parkinsonian syndromes.
• Automated deep learning models for brainstem segmentation require further clinical validation.

Purpose of the Study:

• To develop and validate a two-step deep learning algorithm for brainstem substructure segmentation.
• To classify Parkinsonian syndromes using volumetric measurements derived from segmentation.

Main Methods:

• A retrospective study utilizing internal and external datasets for segmentation and classification.
• Deep learning for segmentation; SVM, random forest, and XGBoost for classification using regional brain volumes.
• Validated using five-fold cross-validation and tested on an external dataset; radiologist performance assessed with and without the model.

Main Results:

• High Dice Similarity Coefficient (DSC) scores for brainstem segmentation (0.969 internal, 0.996 external).
• Support Vector Machine (SVM) achieved the highest classification performance with Area Under the ROC Curve (AUROC) of 0.937 (internal) and 0.914 (external).
• A radiology resident demonstrated improved diagnostic performance when utilizing the model.

Conclusions:

• The developed two-step algorithm effectively combines deep learning and machine learning for automated brainstem segmentation.
• This approach enables reliable differentiation of Parkinsonian syndromes from 3D T1-weighted brain MRI.
• The validated model shows potential for enhancing clinical diagnosis of neurodegenerative diseases.