Deep convolutional-neural-network-based metal artifact reduction for CT-guided interventional oncology procedures...

Area of Science:

• Medical Imaging
• Interventional Radiology
• Artificial Intelligence in Medicine

Background:

• Computed tomography (CT) is crucial for guiding cryoablation but metal artifacts from probes hinder accurate placement and ice ball visualization.
• These artifacts can lead to undertreatment of lesions during CT-guided interventional oncology procedures.
• Improved artifact reduction is needed for enhanced precision in cryoablation.

Purpose of the Study:

• To develop and validate an image-based deep learning model for metal artifact reduction (MARIO) in CT-guided interventional procedures.
• To improve visualization of cryoprobes and ice balls during cryoablation.

Main Methods:

• Developed a novel image domain metal artifact simulation framework for training.
• Generated 361 simulated CT volumes by inserting cryoablation probes into patient scans.
• Trained a U-Net convolutional neural network (CNN) model using simulated data and validated on real cryoablation CT images.

Main Results:

• MARIO processed images showed significantly higher reader scores across all assessed metrics (p < 0.001).
• Improvements observed include: overall image quality (+34.91%), ice ball conspicuity (+36.29%), and needle tip visualization (+34.17%).
• Worst metal artifact reduction and target region confidence showed substantial gains, up to +45.70%.

Conclusions:

• The proposed image-based metal artifact simulation effectively trains a deep learning algorithm (MARIO) for cryoablation.
• MARIO significantly reduces probe-related metal artifacts in CT-guided cryoablation, enhancing procedural accuracy.
• This AI-driven approach shows promise for improving outcomes in interventional oncology.