Deep learning model trained using multi-energy computed tomography (CT) data shows better metal artifact reduction...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computed Tomography

Background:

• Metal artifacts in CT scans, particularly in postoperative lumbar imaging, degrade image quality.
• Virtual Monochromatic Images (VMIs) offer potential for artifact reduction.
• Deep learning models show promise for metal artifact reduction (MAR).

Purpose of the Study:

• To develop and compare deep learning-based MAR (deep-MAR) models using VMIs at single and multiple energy levels.
• To evaluate deep-MAR model performance across a broad spectrum of energy levels (40-140 keV).

Main Methods:

• Developed three deep-MAR models (model70, model100, modelmix) trained on VMIs at 70 keV, 100 keV, and combined levels.
• Utilized multi-energy CT scans from 93 patients with lumbar implants.
• Compared deep-MAR models against a reference (modelMAR) using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), alongside subjective image quality assessments.

Main Results:

• The model trained on mixed energy levels (modelmix) demonstrated comparable or superior PSNR and SSIM compared to single-energy models across 40-140 keV.
• modelmix showed improved attenuation correction performance.
• Deep-MAR models reduced image noise in the spinal canal at 100 keV compared to the reference model.
• Subjective image quality scores for modelmix were comparable or higher than single-energy models.

Conclusions:

• Incorporating multiple energy CT data in deep-MAR model training enhances image quality across broader energy levels.
• This approach is recommended for deep-MAR model development in postoperative lumbar CT scanning.