Synthetic CT generation for pelvic cases based on deep learning in multi-center datasets

Area of Science:

• Medical imaging
• Radiotherapy
• Artificial intelligence

Background:

• Multi-center datasets present challenges for medical image synthesis.
• Rectal cancer radiotherapy often requires both MRI and CT scans.
• Developing an MRI-only workflow is desirable for efficiency and patient comfort.

Purpose of the Study:

• To assess the feasibility of synthesizing CT images from MR images for rectal cancer radiotherapy.
• To develop and validate a generative adversarial network (GAN) model for multi-center MRI-to-CT synthesis.
• To evaluate the dosimetric accuracy of synthetic CT images.

Main Methods:

• A novel GAN model incorporating contrastive learning and consistency regularization was proposed.
• T2-weighted MR and CT images from multi-center datasets were used.
• Image similarity metrics (MAE, SNRpeak, SSIM, GP) and dosimetric analysis were performed.

Main Results:

• The proposed model achieved excellent generalization (GP=0.911) on unseen datasets.
• Significant improvements in image similarity metrics compared to a baseline model were observed.
• Dosimetric analysis showed minimal differences (<1%) between synthetic and real CT images.

Conclusions:

• The developed GAN model accurately synthesizes CT images from MR images in multi-center rectal cancer datasets.
• The synthetic CT images demonstrate dosimetric accuracy within clinically acceptable limits.
• This validates the feasibility of an MRI-only radiotherapy workflow for rectal cancer patients.