Application of 3D-FFDNet technique for Image Denoising in functional MRI
Poster No:
1879
Submission Type:
Abstract Submission
Authors:
Hu Cheng1, Sophia Brink1, Daniel Kennedy1
Institutions:
1Indiana University, Bloomington, IN
First Author:
Co-Author(s):
Introduction:
Deep learning based denoising methods are more robust compared to conventional ones. Among them, FFDnet[1] has emerged as an effective denoising technique based on a convolutional neural network architecture. One of the advantages is it can handle different noise level and has a tuning parameter to balance the blurring and denoising effects. The original FFDNet method was primarily developed for 2D applications. In order to apply this technique in denoising 3D MRI images, we generalized this method to 3D. To our knowledge, it is still a tradition in fMRI data analysis to reduce the noise through spatial smoothing, which decreases the effective resolution and introduce more partial volume effect. Here we propose the application of 3D-FFDNet to reduce the thermal noise in fMRI images without compromising the resolution significantly.
Methods:
Our FFDNet model architecture comprises 5 CNN layers, each with 64 features and a 3-voxel convolution kernel. We trained the model with 61 T1-weight images acquired with the HCP lifespan protocol. Employing a patch size of 44 and a stride of 15, we augmented the dataset, generating 310,000 patches for training. Rician noise was added to the patches for training. The learning rate was initially set to 0.001, and reduced to 0.0001 after 50 epochs. The code was written in pytorch based on the 2D version provided by Tassano et al.[2]. The code is available at https://github.com/huchengMRI.
The 3D-FFDNet denoising was applied on higher resolution MPRAGE images (0.6x0.6x0.8 mm3) and EPI images (1 mm isotropic) acquired on a Siemens 3T scanner. We also applied the denoising method on an HCP dataset (ID: 141826) to investigate the impacts on temporal SNR and brain activation of a finger tapping task. All analyses were conducted using SPM, FSL, AFNI, or Matlab codes.
The 3D-FFDNet denoising was applied on higher resolution MPRAGE images (0.6x0.6x0.8 mm3) and EPI images (1 mm isotropic) acquired on a Siemens 3T scanner. We also applied the denoising method on an HCP dataset (ID: 141826) to investigate the impacts on temporal SNR and brain activation of a finger tapping task. All analyses were conducted using SPM, FSL, AFNI, or Matlab codes.
Results:
Fig. 1A shows the outcome of FFDNet denoising on a high-resolution T1-weighted image, revealing a visually cleaner result that retains intricate anatomical details. In Fig. 1B, the denoising effect on an EPI image commonly used in fMRI is demonstrated, exhibiting enhanced clarity without compromising sharpness. The difference image underscores variations in denoising performance across different brain regions, suggesting potential challenges in preserving contrast in regions with abrupt signal changes. As a comparison, Fig. 1C presents three smoothed images with Gaussian kernels of 2 mm, 4 mm, and 6 mm. As the kernel size increases, the image gets more blurred. The denoised image maintains superior sharpness.
Fig. 2 illustrates the effect of denoising on fMRI activations. First, we conducted a comparative analysis with Gaussian smoothing employing various kernel sizes, quantifying smoothness through AFNI's 3dFWHMx functionality. In Figure 2A, the plot reveals that denoising achieves a level of smoothness equivalent to Gaussian smoothing with a 1.4 mm kernel. However, denoising surpasses Gaussian smoothing in terms of temporal SNR (TSNR) computed on an HCP resting state fMRI dataset (Fig. 2B). Finally, the impact of denoising on fMRI activation during left-hand finger tapping was contrasted with Gaussian smoothing using two different kernels: 1.4 mm and 4 mm. With a 1.4 mm Gaussian kernel, sparse activations were observed in the motor area, whereas the proposed denoising method revealed continuous activation along the gray matter folds. Employing a 4.0 mm Gaussian kernel resulted in continuous activation, but slightly extending into the white matter region.
Fig. 2 illustrates the effect of denoising on fMRI activations. First, we conducted a comparative analysis with Gaussian smoothing employing various kernel sizes, quantifying smoothness through AFNI's 3dFWHMx functionality. In Figure 2A, the plot reveals that denoising achieves a level of smoothness equivalent to Gaussian smoothing with a 1.4 mm kernel. However, denoising surpasses Gaussian smoothing in terms of temporal SNR (TSNR) computed on an HCP resting state fMRI dataset (Fig. 2B). Finally, the impact of denoising on fMRI activation during left-hand finger tapping was contrasted with Gaussian smoothing using two different kernels: 1.4 mm and 4 mm. With a 1.4 mm Gaussian kernel, sparse activations were observed in the motor area, whereas the proposed denoising method revealed continuous activation along the gray matter folds. Employing a 4.0 mm Gaussian kernel resulted in continuous activation, but slightly extending into the white matter region.
Conclusions:
Our study demonstrates the efficacy of 3D-FFDNet denoising in reducing noise while preserving image resolution. Notably, the model, trained on a relatively small dataset of T1-weighted images, exhibits promising noise reduction capabilities when applied to T2-weighted EPI images, surpassing the performance of spatial smoothing in the context of fMRI. Future work will focus on optimizing the model by increasing sample size and including more contrasts.
Modeling and Analysis Methods:
Exploratory Modeling and Artifact Removal 2
Methods Development 1
Keywords:
Data analysis
FUNCTIONAL MRI
1|2Indicates the priority used for review
Provide references using author date format
2. Tassano, M., Delon, J., Veit, T. (2019), An Analysis and Implementation of the FFDNet Image Denoising Method. IPOL, Vol. 9, pp. 1–25. https://doi.org/10.5201/ipol.2019.231.