paper: DP-MDM: Detail-Preserving MR Reconstruction via Multiple Diffusion Models
Authors: Mengxiao Geng, Jiahao Zhu, Ran Hong, Qiqing Liu, Dong Liang, Qiegen Liu*
Physics in Medicine & Biology
https://iopscience.iop.org/article/10.1088/1361-6560/add83a/meta
Date: Feb. 20, 2025
The code and the algorithm are for non-commercial use only.
Copyright 2025, School of Information Engineering, Nanchang University.
Magnetic resonance imaging (MRI) plays a critical role in medical diagnosis and treatment by capturing detailed features, such as subtle tissue changes, which help clinicians make precise diagnoses. However, the widely utilized single diffusion model has limitations, as it fails to accurately capture more intricate details. In this work, we propose a detail-preserving reconstruction method that leverages multiple diffusion models to extract structural and detailed features in the k-space domain, rather than the image domain. Since high-frequency information in k-space is more systematically distributed around the periphery compared to the irregular distribution of detailed information in the image domain, it allows for more efficient extraction of detailed information. To further reduce redundancy and enhance model performance, we introduce virtual binary masks with adjustable circular center windows, which are designed to align with the fre-quency distribution of k-space data, thereby enabling the model to focus attention more efficiently on high-frequency information. Furthermore, the proposed method employs a cascaded architecture, where the first diffusion model mainly recovers low-frequency structural components, and subsequent models enhance high-frequency details during the iterative reconstruction stage. Experimental results demonstrate the effectiveness of the proposed model in enhancing MR reconstruction quality and preserving detailed information.
Pyhton == 3.7.12
jax == 0.3.1
jaxlib == 0.3.0
matplotlib == 3.5.1
ml-collections = 0.1.0
ninja == 1.10.2.3
numpy == 1.21.5
opencv-python == 4.5.5.62
pandas == 1.3.5
Pillow == 9.0.1
protobuf == 3.15.8
scikit-image == 0.16.2
scikit-learn == 1.0.2
scipy == 1.2.0
tensorboard == 2.4.0
tensorflow == 2.4.0
torch == 1.7.0
torchvision == 0.8.0
tqdm == 4.62.3
python main.py --config=./config/ve/SIAT_kdata_ncsnpp.py --workdir=./exp --mode=train --eval_folder=./result
python PCsampling_demo_DPMDM.py
We propose a novel multi-diffusion model framework for the reconstruction of various under-sampled MRI images. Through cascading three diffusion models with gradually expanding masks, the model can learn different levels of details of high-quality K-space images, thereby strengthening the model's understanding of K-space images. The main idea is captured in the figure below:
During the reconstruction phase, in each iteration, the under-sampled image is sequentially passed through three diffusion models. Meanwhile, a data consistency module and relevant masks are added to constrain the generation results of the model. This design enables the model to serve as a unified MRI reconstruction framework that can adapt to various under-sampled MRI.
Overall, we achieved a PSNR of 40.92 and an SSIM of 0.9456 for under-sampled images with 2D Poisson R=10; with 2D Random R=8, we obtained a PSNR of 38.23 and an SSIM of 0.9293; and with Uniform R=10, our approach reached a PSNR of 36.18 and an SSIM of 0.9208 on the SIAT brain MRI dataset. The figure illustrated reconstruction images from MRI images that were under-sampled with 2D Random R=15 and 8 coils, using different models.
