Self-Attention U-Net with Residual Bottleneck for High-Quality CT to MRI Image Translation

Self-Attention U-Net with Residual Bottleneck for High-Quality CT to MRI Image Translation

Stať ve sborníku v databázi WoS či Scopus

The challenge of high-quality CT to MRI image translation remains a critical issue in medical imaging, as MRI scans are expensive and not always accessible, especially in resource-limited settings. This work addresses this problem by presenting a novel Self-Attention U-Net with Residual Bottleneck architecture for generating high-quality MRI images from low-cost CT scans. The model incorporates self-attention mechanisms to capture long-range dependencies, while the residual bottleneck blocks enable efficient feature propagation and preserve fine details during translation. A spectral-normalized Patch Discriminator is employed to improve the realism of generated images, ensuring accurate synthesis of MRI-like structures. Trained on a dataset of 389 paired CT and MRI images, the model is evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), achieving an average PSNR of 23.57 and an average SSIM of 0.6435. These results indicate the model’s ability to generate high-fidelity MRI images from CT scans with competitive perceptual quality. The proposed approach not only shows potential in reducing the reliance on expensive MRI scans but also provides a cost-effective solution for medical imaging, especially in underserved regions. This method could enhance diagnostic capabilities and patient care by enabling the synthesis of MRI images from readily available CT scans, particularly where MRI accessibility is limited.

The challenge of high-quality CT to MRI image translation remains a critical issue in medical imaging, as MRI scans are expensive and not always accessible, especially in resource-limited settings. This work addresses this problem by presenting a novel Self-Attention U-Net with Residual Bottleneck architecture for generating high-quality MRI images from low-cost CT scans. The model incorporates self-attention mechanisms to capture long-range dependencies, while the residual bottleneck blocks enable efficient feature propagation and preserve fine details during translation. A spectral-normalized Patch Discriminator is employed to improve the realism of generated images, ensuring accurate synthesis of MRI-like structures. Trained on a dataset of 389 paired CT and MRI images, the model is evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), achieving an average PSNR of 23.57 and an average SSIM of 0.6435. These results indicate the model’s ability to generate high-fidelity MRI images from CT scans with competitive perceptual quality. The proposed approach not only shows potential in reducing the reliance on expensive MRI scans but also provides a cost-effective solution for medical imaging, especially in underserved regions. This method could enhance diagnostic capabilities and patient care by enabling the synthesis of MRI images from readily available CT scans, particularly where MRI accessibility is limited.

Cross-modal image synthesis; Image translation; GAN Architecture; Medical Imaging; Residual Model; SelfAttention; U-Net

Cross-modal image synthesis; Image translation; GAN Architecture; Medical Imaging; Residual Model; SelfAttention; U-Net

PANDEY, D.; JOSHI, R.; MEZINA, A.; BURGET, R.; DUTTA, M.

2026

03.11.2025

IEEE

Italy

979-8-3315-7675-2

2025 17th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT)

246

251

6