A GAN Guided NCCT to CECT Synthesis With an Advanced CNN-Transformer Aggregated Generator
Computed tomography (CT) is essential for diagnosing and managing various diseases, with contrast-enhanced CT (CECT) offering higher contrast images following contrast agent injection. Nevertheless, the usage of contrast agents may cause side effects. Therefore, achieving high-contrast CT images wit...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10973126/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Computed tomography (CT) is essential for diagnosing and managing various diseases, with contrast-enhanced CT (CECT) offering higher contrast images following contrast agent injection. Nevertheless, the usage of contrast agents may cause side effects. Therefore, achieving high-contrast CT images without the need for contrast agent injection is highly desirable. The main contributions of this paper are as follows: 1) We designed a GAN-guided CNN-Transformer aggregation network called GCTANet for the CECT image synthesis task. We propose a CNN-Transformer Selective Fusion Module (CTSFM) to fully exploit the interaction between local and global information for CECT image synthesis. 2) We propose a two-stage training strategy. We first train a non-contrast CT (NCCT) image synthesis model to deal with the misalignment between NCCT and CECT images. Then we trained GCTANet to predict real CECT images using synthetic NCCT images. 3) A multi-scale Patch hybrid attention block (MSPHAB) was proposed to obtain enhanced feature representations. MSPHAB consists of spatial self-attention and channel self-attention in parallel. We also propose a spatial channel information interaction module (SCIM) to fully fuse the two kinds of self-attention information to obtain a strong representation ability. We evaluated GCTANet on two private datasets and one public dataset. On the neck dataset, the PSNR and SSIM achieved were <inline-formula> <tex-math notation="LaTeX">$35.46\pm 2.783$ </tex-math></inline-formula> dB and <inline-formula> <tex-math notation="LaTeX">$0.970\pm 0.020$ </tex-math></inline-formula>, respectively; on the abdominal dataset, <inline-formula> <tex-math notation="LaTeX">$25.75\pm 5.153$ </tex-math></inline-formula> dB and <inline-formula> <tex-math notation="LaTeX">$0.827\pm 0.073$ </tex-math></inline-formula>, respectively; and on the MRI-CT dataset, <inline-formula> <tex-math notation="LaTeX">$29.61\pm 1.789$ </tex-math></inline-formula> dB and <inline-formula> <tex-math notation="LaTeX">$0.917\pm 0.032$ </tex-math></inline-formula>, respectively. In particular, in the area around the heart, where obvious movements and disturbances were unavoidable due to the heartbeat and breathing, GCTANet still successfully synthesized high-contrast coronary arteries, demonstrating its potential for assisting in coronary artery disease diagnosis. The results demonstrate that GCTANet outperforms existing methods. |
|---|---|
| ISSN: | 2169-3536 |