Magnetic Resonance-Based Synthetic Computed Tomography Using Generative Adversarial Networks for Intracranial Tumor...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Computed Tomography (CT) is crucial for radiotherapy planning, but Magnetic Resonance Imaging (MRI) offers superior soft-tissue contrast.
• Synthesizing CT from MRI can bridge this gap, improving treatment accuracy and patient outcomes.

Purpose of the Study:

• To develop a reliable deep learning method for synthesizing CT from MRI, enhancing resolution for radiotherapy planning.
• To evaluate the accuracy and clinical utility of the synthesized CT (sCT) images.

Main Methods:

• A 3D conditional generative adversarial network (pix2pix) was employed to map MRI data to CT data.
• The model was trained on 26 paired MRI-CT datasets and validated on 5.
• Quantitative similarity metrics (cosine angle distance, Euclidean distance, MSE, PSNR, MSSIM) and qualitative radiologist evaluations were used.

Main Results:

• High similarity indices were achieved: MSSIM of 0.84 ± 0.036, PSNR of 29.47 ± 1.35.
• Radiologists reported excellent satisfaction with spatial geometry and noise levels, good satisfaction with contrast and artifacts, and fair imaging details.
• Quantitative and qualitative assessments confirmed the usability of the synthetic CT.

Conclusions:

• The developed deep learning method reliably synthesizes high-resolution CT from MRI for radiotherapy planning.
• The sCT images demonstrate significant potential for clinical application, offering a viable alternative where CT is unavailable or suboptimal.
• Further refinement could improve imaging detail, enhancing its utility in complex treatment scenarios.