Influence of Image Resolution on GAN Models for MRI Super-Resolution
Poster No:
1625
Submission Type:
Abstract Submission
Authors:
Chen-Chia Hsu1, Yi-Ping Chao2, Li-Wei Kuo3, Kuan-Hung Cho1
Institutions:
1Department of Electronic Engineering, National United University, Miaoli, Taiwan, 2Department of Computer Science and Information Engineering, Chang Gung University, Taoyuan City, Taiwan, 3Institute of Biomedical Engineering and Nanomedicine, National Health Research Institutes, Miaoli, Taiwan
First Author:
Co-Author(s):
Yi-Ping Chao
Department of Computer Science and Information Engineering, Chang Gung University
Taoyuan City, Taiwan
Department of Computer Science and Information Engineering, Chang Gung University
Taoyuan City, Taiwan
Li-Wei Kuo
Institute of Biomedical Engineering and Nanomedicine, National Health Research Institutes
Miaoli, Taiwan
Institute of Biomedical Engineering and Nanomedicine, National Health Research Institutes
Miaoli, Taiwan
Introduction:
MRI is a vital diagnostic imaging modality capable of reconstructing images across multiple contrasts. Its versatility enables MRI to provide detailed insights into anatomical, physiological, cellular, and molecular characteristics, making it indispensable for both preclinical and clinical studies.
In recent years, the advent of neural networks has led to the widespread adoption of deep learning (DL)-based super-resolution (SR) techniques to enhance MRI image resolution (Altmann, 2024; Chatterjee, 2021; Hwang 2024; Lee, 2024; Liu 2021) These methods allow the restoration of HR details crucial for clinical diagnosis without requiring modifications to existing systems or hardware, offering a cost-effective solution with reduced scanning times. Despite their potential, the impact of input image resolution on the performance of DL-based SR remains largely unexplored.
This study investigates the performance of Super-Resolution Generative Adversarial Networks (SRGANs) on human brain MRI images with varying resolutions. Our findings reveal that low-resolution (LR) images with larger voxel sizes result in poorer reconstruction of HR images using DL-based SR methods. This highlights the importance of input resolution in optimizing the performance of SR techniques for clinical applications.
In recent years, the advent of neural networks has led to the widespread adoption of deep learning (DL)-based super-resolution (SR) techniques to enhance MRI image resolution (Altmann, 2024; Chatterjee, 2021; Hwang 2024; Lee, 2024; Liu 2021) These methods allow the restoration of HR details crucial for clinical diagnosis without requiring modifications to existing systems or hardware, offering a cost-effective solution with reduced scanning times. Despite their potential, the impact of input image resolution on the performance of DL-based SR remains largely unexplored.
This study investigates the performance of Super-Resolution Generative Adversarial Networks (SRGANs) on human brain MRI images with varying resolutions. Our findings reveal that low-resolution (LR) images with larger voxel sizes result in poorer reconstruction of HR images using DL-based SR methods. This highlights the importance of input resolution in optimizing the performance of SR techniques for clinical applications.
Methods:
A total of 120 subjects' 3D T1-weighted images (T1WI) with a voxel size of 0.7 mm isotropic were obtained from the Human Connectome Project 1200 Subjects Data, while another set of 120 subjects' 3D T1WI with a voxel size of 1 mm isotropic was sourced from the OASIS dataset provided by the Washington University School of Medicine. These HR images served as the reference data.
LR images were generated by downscaling the HR images by factors of 2 and 4. This process yielded voxel sizes of 1.4 mm and 2.8 mm from the 0.7 mm dataset, and 2 mm and 4 mm from the 1 mm dataset. SR images were then generated using SRGAN with 2-fold scaling for images with voxel sizes of 1.4 mm, 2 mm, 2.8 mm, and 4 mm, and these were compared to HR images with voxel sizes of 0.7 mm, 1 mm, 1.4 mm, and 2 mm, respectively. Additionally, SR images with 4-fold scaling from voxel sizes of 2.8 mm and 4 mm were compared to HR images with voxel sizes of 0.7 mm and 1 mm, respectively.
The SRGAN model was trained for 100 epochs, leveraging adversarial learning to generate optimized SR images. The quality of the SR images was assessed by comparing them to the original HR images using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) as evaluation metrics.
LR images were generated by downscaling the HR images by factors of 2 and 4. This process yielded voxel sizes of 1.4 mm and 2.8 mm from the 0.7 mm dataset, and 2 mm and 4 mm from the 1 mm dataset. SR images were then generated using SRGAN with 2-fold scaling for images with voxel sizes of 1.4 mm, 2 mm, 2.8 mm, and 4 mm, and these were compared to HR images with voxel sizes of 0.7 mm, 1 mm, 1.4 mm, and 2 mm, respectively. Additionally, SR images with 4-fold scaling from voxel sizes of 2.8 mm and 4 mm were compared to HR images with voxel sizes of 0.7 mm and 1 mm, respectively.
The SRGAN model was trained for 100 epochs, leveraging adversarial learning to generate optimized SR images. The quality of the SR images was assessed by comparing them to the original HR images using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) as evaluation metrics.
Results:
Figure 1 illustrates the reconstructed SR images generated from different voxel sizes and scaling factors. For 2-fold SR images, those reconstructed from 2.8 mm and 4 mm voxel sizes appear slightly noisier compared to those reconstructed from 1.4 mm and 2 mm voxel sizes, respectively. Additionally, while 2-fold SR images from 2.8 mm and 4 mm LR images are noisier, they exhibit greater clarity compared to 4-fold SR images derived from the same voxel sizes.
Table 1 provides the PSNR and SSIM metrics for each scenario. The results indicate that the PSNR decreases significantly at a voxel size of 4 mm for both 2-fold and 4-fold scaling. Similarly, the SSIM index shows a slight decline as the voxel size increases. Furthermore, for voxel sizes of 2.8 mm and 4 mm, the SSIM index decreases noticeably with higher scaling factors, highlighting the diminishing reconstruction quality with increased folds.
Table 1 provides the PSNR and SSIM metrics for each scenario. The results indicate that the PSNR decreases significantly at a voxel size of 4 mm for both 2-fold and 4-fold scaling. Similarly, the SSIM index shows a slight decline as the voxel size increases. Furthermore, for voxel sizes of 2.8 mm and 4 mm, the SSIM index decreases noticeably with higher scaling factors, highlighting the diminishing reconstruction quality with increased folds.
·Figure 1. The results of SR images from LR images with different voxel size
Conclusions:
In this study, we generate LR images from HR images to evaluate the influence from the MRI voxel size on the performance of SR technique. Our results show that the PSNR and SSIM indices will decrease significantly as the voxel size greater than 2.8 mm, which suggests that the performance of DL-based SR reconstruction on MRI images will be influenced by the voxel size. According to our results, it is better to apply DL-based SR techniques on MRI images with voxel size smaller than about 3 mm.
Modeling and Analysis Methods:
Other Methods 1
Novel Imaging Acquisition Methods:
Anatomical MRI 2
Keywords:
Other - Super resolution
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
Provide references using APA citation style.
2. Chatterjee, S., Sciarra, A., Dünnwald, M., Mushunuri, R. V., Podishetti, R., Rao, R. N., Gopinath, G. D., Oeltze-Jafra, S., Speck, O., & Nürnberger, A. (2021, 23-27 Aug. 2021). ShuffleUNet: Super resolution of diffusion-weighted MRIs using deep learning. 2021 29th European Signal Processing Conference (EUSIPCO).
3. Hwang, J. H., Park, C. K., Kang, S. B., Choi, M. K., & Lee, W. H. (2024). Deep Learning Super-Resolution Technique Based on Magnetic Resonance Imaging for Application of Image-Guided Diagnosis and Surgery of Trigeminal Neuralgia. Life (Basel), 14(3).
4. Lee, J., Jung, W., Yang, S., Park, J. H., Hwang, I., Chung, J. W., Choi, S. H., & Choi, K. S. (2024). Deep learning-based super-resolution and denoising algorithm improves reliability of dynamic contrast-enhanced MRI in diffuse glioma. Sci Rep, 14(1), 25349.
5. c, H., Liu, J., Li, J., Pan, J. S., & Yu, X. (2021). DL-MRI: A Unified Framework of Deep Learning-Based MRI Super Resolution. J Healthc Eng, 2021, 5594649.
6. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., & Shi, W. (2017, 21-26 July 2017). Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),