Curriculum Based Multi-Task Learning for Parkinson's Disease Detection.

Area of Science:

• Neuroimaging
• Machine Learning
• Neurology

Background:

• Parkinson's disease (PD) diagnosis is challenging in early stages, necessitating advanced detection methods.
• Radiological classifiers are crucial for PD diagnosis, staging, and predictive modeling.
• Deep learning, particularly CNNs, shows promise for analyzing medical images.

Approach:

• A curriculum learning strategy was developed to train a CNN using severity-based metadata from the Hoehn and Yahr (H&Y) staging system for PD.
• The training data was progressively increased in difficulty, starting with easier-to-classify samples.
• Multi-task learning with pre-trained CNNs and transfer learning were employed for PD classification using T1-weighted (T1-w) MRI scans.

Key Points:

• Curriculum training significantly boosted PD classification performance by 3.9% compared to the baseline model.
• The study utilized a dataset of 1,012 participants (653 PD patients, 359 controls) aged 20.0-84.9 years.
• Despite challenges, T1-w MRI classification achieved an ROC AUC of 0.59-0.65, with improvement via curriculum learning.

Conclusions:

• Curriculum learning enhances the performance of CNN-based classifiers for Parkinson's disease detection from MRI.
• This approach offers a promising strategy for improving early diagnosis and management of neurodegenerative diseases.
• Future research incorporating multimodal imaging may further improve classification accuracy.