Incorporating Radiologist Knowledge Into MRI Quality Metrics for Machine Learning Using Rank-Based Ratings

Area of Science:

• Medical Imaging
• Deep Learning
• Radiology

Background:

• Traditional image quality metrics are inadequate for medical images.
• Deep learning models require specialized metrics for optimal performance.

Purpose of the Study:

• To develop a novel, MRI-specific image quality metric using radiologist rankings and deep learning.
• To evaluate the performance of this metric in optimizing deep learning tasks for medical imaging.

Main Methods:

• A retrospective analysis of 19,344 MRI image pair rankings from the NYU fastMRI Initiative neuro database.
• Training deep learning models (EfficientNet, IQ-Net) on radiologist rankings, comparing them with Mean Squared Error (MSE) and Structural Similarity (SSIM).
• Evaluating DL models optimized with the new metric for denoising and reconstruction tasks.

Main Results:

• Radiologist image ranking showed higher agreement (70.4%) than Likert scoring (25%).
• Deep learning models, particularly IQ-Net, accurately predicted radiologist rankings.
• IQ-Net optimized networks demonstrated superior performance in both denoising and reconstruction tasks compared to EfficientNet and traditional metrics.

Conclusions:

• Deep learning models trained on radiologist image rankings can serve as effective image quality metrics.
• The developed IQ-Net metric successfully optimizes deep learning tasks in MRI, improving image quality.