Automatic MR image quality evaluation using a Deep CNN: A reference-free method to rate motion artifacts in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Motion artifacts significantly degrade magnetic resonance (MR) image quality, impacting clinical diagnosis and research.
• Identifying these artifacts is crucial but challenging due to the difficulty in preventing patient motion during scans.

Purpose of the Study:

• To develop an automated method for detecting motion artifacts in MR images using deep convolutional neural networks (CNNs).
• To leverage transfer learning and fine-tuning to overcome the scarcity of annotated medical imaging data.

Main Methods:

• Four pre-trained CNN architectures were fine-tuned using annotated MR image patches.
• Models were trained and tested on T1-weighted volumetric acquisitions from healthy volunteers.
• A consensus approach combining shallower architectures and an artificial neural network (ANN) classifier was used for final detection.

Main Results:

• The fine-tuned models achieved an average accuracy of 86.3% for detecting artifacts on MR image patches across three axes.
• Exploring shallower network layers improved patch-level accuracy to 90.4%.
• The final ANN classifier combining shallower models achieved 100.0% accuracy in identifying motion-corrupted MR acquisitions.

Conclusions:

• The proposed automated method effectively detects motion artifacts in MR images with high accuracy.
• While effective, generalization to diverse datasets (e.g., epilepsy patients, different vendors) indicates a need for domain adaptation.
• The method provides a rapid, scalable solution for quality control in large MR image datasets.