CT synthesis from CBCT using a sequence-aware contrastive generative network.

Area of Science:

• Medical Imaging
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Synthetic CT (sCT) generation from Cone-Beam CT (CBCT) is vital for adaptive radiotherapy, enabling timely dose calculations and plan adjustments.
• Existing Cycle-GAN methods for sCT synthesis face limitations, including unmet bijective assumptions and failure to leverage multi-set CBCT data.

Purpose of the Study:

• To propose a novel framework, the sequence-aware contrastive generative network (SCGN), to enhance CBCT quality for improved sCT synthesis.
• To address the limitations of Cycle-GAN in CBCT to CT conversion by incorporating attention mechanisms and contrastive learning.

Main Methods:

• Developed an attention sequence fusion module to improve CBCT image quality.
• Applied contrastive learning within generative adversarial networks (GANs) to focus on anatomical structures in CBCT feature extraction.
• Introduced a new generator architecture to enhance the accuracy of anatomical details in synthetic CT images.

Main Results:

• The proposed SCGN framework demonstrated significant improvements in unsupervised CT synthesis.
• Experimental results confirmed the superior performance of SCGN compared to existing unsupervised CT synthesis methods on the tested datasets.
• The method effectively utilizes complementary information from multiple CBCT sets for enhanced synthesis.

Conclusions:

• The SCGN framework offers a promising solution for high-quality synthetic CT generation from CBCT data.
• This advancement has the potential to improve the efficiency and accuracy of adaptive radiotherapy planning.
• The study highlights the effectiveness of integrating attention mechanisms and contrastive learning for medical image synthesis.