Unsupervised Simultaneous Denoising and Cross-Modality Synthesis of Medical Image via Generative Adversarial Networks

Multi-modality medical imaging provides complementary anatomical and
functional information that is essential for accurate diagnosis and treatment
planning. In clinical practice, however, the acquisition of multiple high-
quality imaging modalities is often constrained by long scanning times, high
financial cost, and patient discomfort. Medical images are also frequently de-
graded by noise introduced during acquisition, particularly under low-dose
or high-speed scanning conditions. Although denoising and cross-modality
synthesis are closely related in real-world workflows, existing methods com-
monly treat them as independent problems.
This thesis presents an unsupervised unified framework for the simul-
taneous denoising of medical images and cross-modality image synthesis. A
Multi-Channel Asymmetric Residual Generator (MARG) is proposed to en-
hance noise suppression while preserving fine anatomical structures. In ad-
dition, a Dual-Channel Joint Discriminator (DCD) is developed to improve
modality translation by jointly enforcing local structural consistency and pixel-
level intensity distribution. These components are integrated into a single
adversarial learning framework that enables concurrent denoising and cross-
modality synthesis.
Experiments conducted on low-dose chest CT and multi-modal brain MRI
datasets demonstrate that the proposed methods outperform state-of-the-art
unsupervised baselines in terms of PSNR, SSIM, and MS-SSIM. For simul-
taneous denoising and synthesis tasks, the proposed framework achieves
improved image fidelity and structural consistency compared with conven-
tional sequential processing strategies across multiple noise levels.
By reducing reliance on paired training data and mitigating error accu-
mulation associated with multi-stage pipelines, the proposed approach of-
fers a practical and scalable solution for enhancing image quality and mod-
ality availability in clinical environments. The framework has the potential
to improve diagnostic reliability , workflow efficiency , and overall patient ex-
perience.