Deep learning-based virtual noncalcium imaging in multiple myeloma using dual-energy CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Dual-energy CT with virtual noncalcium (VNCa) images aids in evaluating bone marrow involvement in multiple myeloma.
• Current VNCa techniques face challenges with image noise and artifacts due to material decomposition.

Purpose of the Study:

• To enhance VNCa image quality for multiple myeloma assessment using an AI-based method called AGATE.
• To improve the visualization of focal intramedullary bone marrow lesions.

Main Methods:

• Developed a dual-task convolutional neural network (CNN) for material classification and quantification.
• Optimized CNN parameters using custom loss functions incorporating physics-informed constraints and texture information.
• Trained the model on CT phantoms and validated generalizability on phantom and patient data, assessing material decomposition accuracy, noise power spectrum (NPS), and modulation transfer function (MTF).

Main Results:

• AGATE achieved consistent material quantification accuracy (MAPE 0.7%-4.4%) across various phantom sizes and radiation doses.
• AGATE-synthesized VNCa images showed 50-77% noise reduction compared to commercial VNCa images, without compromising spatial resolution.
• Patient VNCa images generated by AGATE exhibited reduced noise and artifacts, leading to improved delineation of multiple myeloma lesions.

Conclusions:

• The AGATE method effectively reduces noise and artifacts in VNCa images.
• AGATE enhances the visualization of bone marrow lesions, improving diagnostic capabilities for multiple myeloma assessment.