Probabilistic Brain Extraction in MR Images via Conditional Generative Adversarial Networks.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Brain extraction is crucial for neuroimaging applications like volume quantification and disease monitoring.
• Existing deep learning methods improve accuracy but do not address prediction uncertainty.
• Quantifying uncertainty in brain extraction is essential for reliable downstream analysis.

Purpose of the Study:

• To develop a novel probabilistic deep learning algorithm for brain extraction that quantifies uncertainty.
• To recast brain extraction as a Bayesian inference problem solvable with a conditional generative adversarial network (cGAN).

Main Methods:

• Utilized a cGAN where the generator takes an MR head image as input and outputs multiple plausible brain images.
• Generated a pixel-wise mean image for brain extraction and a standard deviation image to quantify prediction uncertainty.
• Tested the algorithm on diverse datasets (NFBS, CC359, LPBA, IBSR) with heterogeneous imaging factors.

Main Results:

• The proposed probabilistic method demonstrated superior accuracy and robustness compared to a standard tool.
• Achieved accuracy comparable to existing deep learning methods while additionally providing uncertainty quantification.
• Highlighted the value of uncertainty quantification in neuroimaging applications.

Conclusions:

• The novel probabilistic deep learning approach effectively performs brain extraction and quantifies prediction uncertainty.
• This method offers a more robust and informative alternative to existing brain extraction techniques.
• Quantifying uncertainty is vital for advancing the reliability and interpretability of neuroimaging studies.