Deep learning, data ramping, and uncertainty estimation for detecting artifacts in large, imbalanced databases of MRI...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Magnetic resonance imaging (MRI) is crucial for identifying neurological disorder-related changes.
• Current MRI quality assessment relies on manual inspection, which is time-consuming and prone to variability.
• Automated artifact detection is needed for efficient, reliable analysis of large neuroimaging datasets.

Purpose of the Study:

• To develop a high-performing, automated MRI artifact detection algorithm for large and imbalanced datasets.
• To incorporate uncertainty quantification into deep learning models for MRI artifact detection.
• To address the challenge of class imbalance in neuroimaging data.

Main Methods:

• Implemented a stochastic deep learning algorithm using Monte Carlo dropout in a 3D AlexNet.
• Utilized data-ramping for transfer learning to manage class imbalance in a large dataset (34,800 scans).
• Estimated both epistemic and aleatoric uncertainties to improve detection reliability.

Main Results:

• Achieved 94.9% testing accuracy with data-ramping, outperforming focal cross-entropy.
• Significantly improved testing accuracy to 99.5% by implementing epistemic uncertainties.
• Demonstrated the feasibility of incorporating aleatoric uncertainties into the detection pipeline.

Conclusions:

• The proposed stochastic deep learning approach offers a robust solution for automated MRI artifact detection in large, imbalanced datasets.
• Incorporating epistemic and aleatoric uncertainties enhances the reliability and accuracy of artifact detection.
• This method improves the efficiency of managing neuroimaging databases and excluding artifact-contaminated data.