DGCD-3D: Difference-guided conditional diffusion model for low-field 3D MRI enhancement to assist stroke assessment.
Authors
Affiliations (7)
Affiliations (7)
- School of Biological Science and Medical Engineering, Beihang University, Beijing 100191, China.
- Beijing Celebrain Technology Co., Ltd., Beijing 100083, China.
- China National Clinical Research Center for Neurological Diseases, Department of Neurology, Beijing Tiantan Hospital, Capital Medical University, Beijing 100070, China.
- China National Clinical Research Center for Neurological Diseases, Beijing Tiantan Hospital, Capital Medical University, Beijing 100070, China.
- School of Biological Science and Medical Engineering, Beihang University, Beijing 100191, China. Electronic address: [email protected].
- China National Clinical Research Center for Neurological Diseases, Department of Neurology, Beijing Tiantan Hospital, Capital Medical University, Beijing 100070, China. Electronic address: [email protected].
- School of Biological Science and Medical Engineering, Beihang University, Beijing 100191, China. Electronic address: [email protected].
Abstract
Low-field (LF) magnetic resonance imaging (MRI) plays a crucial role in assisting clinicians with rapid stroke diagnosis. However, its inherent limitations, such as low signal-to-noise ratio (SNR) and suboptimal image quality, make accurate stroke lesion identification challenging. To address this, we propose a Difference-Guided Conditional Diffusion Model (DGCD-3D) to enhance image quality of LF MRI while preserving structural integrity of stroke lesions. Specifically, the model incorporates a difference-adaptive forward diffusion process that guides the diffusion dynamics based on differences. During training, multi-scale intrinsic features from LF MRI and prior spatial information of stroke lesions are explicitly encoded into the generative process. Furthermore, a time-adaptive multi-loss optimization strategy dynamically balances pixel-wise and perceptual losses at different timesteps. DGCD-3D was evaluated on a large-scale, clinically scarce paired LF-HF MRI dataset (n = 974) acquired within a mean interval of 18.5 min. Experimental results demonstrate that DGCD-3D substantially improves LF MRI image quality (PSNR = 28.26, SSIM = 0.896) and achieves significantly higher consistency with HF MRI in stroke lesion assessment (Spearman's correlation coefficient ρ = 0.732) compared with the LF MRI (ρ = 0.680). Furthermore, a clinical authenticity assessment conducted by six experienced radiologists yielded a confusion rate of 51.6% and a confusion score of 5.64, further confirming the clinical reliability and broad applicability of the proposed approach. The codes and trained models will be released on GitHub: https://github.com/lihao9056/DGCD-3D.