Deep Convolutional Neural Networks on Multiclass Classification of Three-Dimensional Brain Images for Parkinson's...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Parkinson's disease (PD) is a neurodegenerative disorder impacting motor function.
• Accurate staging of PD is crucial for effective treatment and management.
• Functional medical imaging, particularly single-photon emission computed tomography (SPECT), is vital for PD diagnosis.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting Parkinson's disease stages using 3D brain SPECT images.
• To compare the performance of 2D and 3D Convolutional Neural Networks (CNNs) with and without attention mechanisms.
• To investigate the utility of cotraining for enhancing model robustness across multi-center datasets.

Main Methods:

• Utilized two multi-center 3D SPECT datasets (n=634, n=202) for Parkinson's disease staging.
• Implemented 2D CNNs (pretrained on ImageNet) and 3D CNNs (pretrained on Kinetics-400).
• Incorporated an attention mechanism and employed cotraining for simultaneous multi-dataset training.

Main Results:

• 2D CNNs pretrained on ImageNet outperformed 3D CNNs pretrained on Kinetics-400.
• Models incorporating an attention mechanism demonstrated superior performance over standard 2D and 3D CNNs.
• Cotraining effectively improved model performance with sufficiently large datasets.

Conclusions:

• Deep learning models, particularly 2D CNNs with attention, show significant promise for automated Parkinson's disease staging using 3D SPECT.
• Attention mechanisms enhance the model's ability to focus on relevant image features for improved diagnostic accuracy.
• Multi-center training strategies like cotraining can improve the generalizability and robustness of AI models in clinical neuroimaging.