Convolutional Neural Networks for Neuroimaging in Parkinson's Disease: Is Preprocessing Needed?

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Spatial and intensity normalizations are standard but computationally expensive preprocessing steps in neuroimaging analysis.
• These normalization techniques can introduce image deformations, potentially affecting diagnostic information.
• Nuclear imaging, like PET-FDG and FP-CIT SPECT, is vital for conditions such as Parkinson's disease.

Purpose of the Study:

• To investigate the capability of Convolutional Neural Networks (CNNs) to handle spatial and intensity variations in nuclear brain imaging.
• To determine if traditional spatial and intensity normalization preprocessing steps are still necessary when using CNNs for analysis.

Main Methods:

• Four CNN models with established architectures were trained, with and without spatial and intensity normalization preprocessing.
• A 3D version of the ALEXNET architecture was utilized for its complexity and effectiveness.

Main Results:

• A complex CNN model, ALEXNET, demonstrated the ability to account for spatial differences without explicit normalization, achieving 94.1% diagnostic accuracy and an ROC AUC of 0.984.
• Saliency map visualizations confirmed that the CNN models identified relevant patterns consistent with existing literature.
• Intensity normalization was found to significantly impact model performance and accuracy, underscoring its importance.

Conclusions:

• Sophisticated CNNs can effectively mitigate the need for spatial normalization in nuclear brain imaging analysis.
• Intensity normalization remains a critical factor influencing the accuracy and reliability of CNN-based diagnostic models.
• Future research should focus on optimizing intensity normalization strategies for CNN applications in neuroimaging.