Deep learning encodes robust discriminative neuroimaging representations to outperform standard machine learning.

Area of Science:

• Neuroimaging and Machine Learning
• Computational Neuroscience
• Artificial Intelligence in Medical Imaging

Background:

• Critical commentaries often compare deep learning (DL) and standard machine learning (SML) for brain imaging, but may overlook DL's representation learning capabilities.
• Previous comparisons may have used pre-engineered features, limiting DL's potential by not leveraging its core strength.
• Standard machine learning (SML) methods have been the benchmark, but their efficacy with complex neuroimaging data is being re-evaluated.

Purpose of the Study:

• To conduct a large-scale, systematic comparison of deep learning (DL) versus standard machine learning (SML) for structural MRI data analysis.
• To demonstrate the critical importance of representation learning for DL's performance in neuroimaging tasks.
• To evaluate DL's scalability, computational complexity, and ability to identify task-discriminative biomarkers in brain imaging.

Main Methods:

• A large-scale, systematic comparison was performed on multiple classification and regression tasks using structural MRI images.
• Deep learning (DL) models were trained following prevalent DL practices, emphasizing end-to-end representation learning.
• Performance was benchmarked against standard machine learning (SML) methods, analyzing scalability, computational time, and biomarker localization.

Main Results:

• Deep learning (DL) methods, when utilizing representation learning, show substantial improvement over standard machine learning (SML) approaches.
• DL methods exhibit favorable scalability and lower asymptotic computational complexity relative to SML, despite inherent complexity.
• DL embeddings consistently identify task-discriminative brain biomarkers and reveal comprehensible, task-specific projection spectra.

Conclusions:

• Representation learning is crucial for unlocking the full potential of deep learning (DL) in brain imaging analysis.
• Deep learning (DL) effectively exploits nonlinearities in neuroimaging data to generate superior representations for brain characterization.
• DL offers a powerful approach for identifying brain biomarkers and advancing the analysis of structural MRI data compared to SML.