Synthetic CT generation based on multi-sequence MR using CycleGAN for head and neck MRI-only planning

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy Physics

Background:

• Accurate computed tomography (CT) imaging is crucial for radiotherapy planning in nasopharyngeal carcinoma (NPC).
• Magnetic resonance (MR) imaging offers superior soft-tissue contrast but lacks electron density information for direct use in dose calculations.
• Synthetic CT (sCT) generation from MR data aims to bridge this gap, but sequence selection impacts accuracy.

Purpose of the Study:

• To evaluate the influence of various MR sequences (T1, T2, T1C, T1DIXONC) on the accuracy of CycleGAN-generated sCT for NPC.
• To compare the performance of single-sequence versus multi-sequence MR-based sCT generation.
• To assess the dosimetric feasibility of the generated sCT images.

Main Methods:

• Utilized a dataset of 143 head and neck MR (T1, T2, T1C, T1DIXONC) and CT scans from NPC patients.
• Improved CycleGAN architecture with enhanced generator/discriminator and a novel cyclic consistent structure control domain loss function.
• Evaluated sCT accuracy using quantitative metrics (MAE, PSNR, SSIM, RMSE) and metrological evaluation (gamma analysis at 3%/3 mm).

Main Results:

• T1 sequence-based sCT demonstrated superior accuracy compared to other single MR sequences.
• Multi-sequence MR-based sCT, especially incorporating T1, outperformed single-sequence approaches across evaluation metrics.
• Global gamma passage rates exceeded 95% (3%/3 mm) for most MR-based sCT, indicating good dosimetric agreement, except for the T2 sequence.

Conclusions:

• CycleGAN effectively synthesizes CT images from MR data for NPC patients.
• T1-weighted MR sequences are most effective for single-sequence sCT generation.
• Multi-sequence MR input, leveraging T1, provides the highest accuracy and dosimetric quality for sCT, showing significant potential for radiotherapy applications.