A Deep Learning System for Synthetic Knee Magnetic Resonance Imaging: Is Artificial Intelligence-Based Fat-Suppressed...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Orthopedic Imaging

Background:

• Fat-suppressed (FS) MRI is crucial for visualizing knee structures but can be time-consuming.
• Deep learning offers potential for generating synthetic MRI sequences to improve efficiency.

Purpose of the Study:

• To evaluate the diagnostic performance and quality of artificial intelligence-based FS MRI (AFSMRI) for knee abnormalities.
• To assess the interchangeability of AFSMRI with original MRI protocols.

Main Methods:

• Synthetic AFSMRI scans were generated using a deep learning U-Net model.
• Three radiologists assessed 88 knee MR studies comparing original (PD + FSPD) and synthetic (PD + AFSMRI) sequences.
• Diagnostic performance, image quality, intermethod/interobserver agreement, and contrast-to-noise ratio were analyzed.

Main Results:

• AFSMRI quality was rated as diagnostic in 98.9%-100% of cases.
• High sensitivity and specificity were observed for meniscal, cruciate ligament, collateral ligament, and tendon tears.
• Intermethod agreement was moderate to substantial, and original and synthetic protocols were interchangeable for diagnosing knee abnormalities.

Conclusions:

• Artificial intelligence-based FS MRI (3D AFSMRI) is a feasible and fast imaging method.
• AFSMRI demonstrates comparable detection rates for knee structural abnormalities to original 3D MR sequences.
• This AI-driven approach offers a promising alternative for efficient knee MRI.