Artifact correction for functional mri using overfitted neural network: a feasibility study

2326 

Abstract Submission 

Soo-hyung Lee¹, Kyu-Jin Jung², Chuanjiang Cui³, Hyun-joo Song⁴, Ji-Won Chun⁵, Dong-Hyun Kim³

¹Department of Electrical & Electronic Engineering, Yonsei University, Seoul, Korea, ²Department of Electrical & Electronic Engineering, Yonsei University, Seoul, Korea, Republic of, ³Department of Electrical & Electronic Engineering, Yonsei University, Seoul, Not Applicable, ⁴Department of Psychology, Yonsei University, Seould, Korea, Republic of, ⁵Catholic University of Korea College of Medicine, Seoul, Not Applicable

Soo-hyung Lee  
Department of Electrical & Electronic Engineering, Yonsei University
Seoul, Korea

Kyu-Jin Jung  
Department of Electrical & Electronic Engineering, Yonsei University
Seoul, Korea, Republic of

Chuanjiang Cui  
Department of Electrical & Electronic Engineering, Yonsei University
Seoul, Not Applicable

Hyun-joo Song  
Department of Psychology, Yonsei University
Seould, Korea, Republic of

Ji-Won Chun  
Catholic University of Korea College of Medicine
Seoul, Not Applicable

Dong-Hyun Kim  
Department of Electrical & Electronic Engineering, Yonsei University
Seoul, Not Applicable

Functional magnetic resonance imaging (fMRI) typically uses Gradient Echo EPI (GE-EPI) to detect the blood oxygenation level-dependent (BOLD) signal. GE-EPI offers high BOLD sensitivity but struggles with artifacts such as B0 inhomogeneity and geometric distortion, limiting observations in regions of the frontal or auditory areas. Spin Echo(SE-EPI) counteracts inhomogeneity artifact-induced signal dropout and use of reversed acquisition in the readout direction (so-called top-up correction [1]) can also reduce the geometric distortion. However, SE-EPI has reduced BOLD sensitivity. Motivated by the above properties, we propose a neural network pipeline which reduces the inhomogeneity artifact of GE-EPI images. The proposed method leverages overfitting of neural network by training the GE-EPI onto a single additional data set of the SE-EPI acquisition.

Fig.1 presents the pipeline of the algorithm. The trainset of the network consists of GE-EPI and top-up corrected SE-EPI [1] as label from individual single subject. Data were acquired using Siemens 3T scanner. An actual fMRI task was performed using GE-EPI. For the GE-EPI TE=30ms, TR=3000ms, bandwidth=2894 KHz, resolution=(1.9,1.9,5)mm, matrix size=(128,128,33), # of measurements=90. For SE-EPI TE=30ms, TR=3000ms, bandwidth=2894 KHz, resolution = (1.9,1.9,5)mm, matrix size=(128,128,33), # of measurements = 2 for a scan time of 30 sec. The pipeline starts with GE-EPI (during event-OFF period) and SE-EPI data. GE-EPI were averaged to one 3D volume to increase SNR level. According to prior work that signal dropout and distortion correction can be seen as a deconvolution problem [2], we assumed the correction as a spectrum domain deconvolution operation. Hence, the averaged GE-EPI and corrected SE-EPI were Fourier transformed and subsequently used for train input and label for the network. A network based on the ResNet structure was used [3]. Fc layer and bias were removed. Different number of epochs and layer features were tested to control the level of overfitting. After training, the model was applied to the original GE-EPI data (both event-ON and event-OFF period). A visual stimulation experiment was performed to compare the BOLD activation after the correction of the proposed method.

Fig. 2 show results of the proposed method. The network's reconstruction can be seen in Fig. 2(a). As seen, the corrected GE-EPI images from the output of the network shows high similarity with the label images. While the structural reconstruction of the network shows high SSIM values, this does not guarantee the ability to retrieve the BOLD activation. Fig2 (b) shows BOLD results from the visual stimulation experiment. By observing the BOLD activation maps with different hyperparameter settings of the network, we observed that network had trade-off between structural reconstruction ability and BOLD reconstruction ability which were dependent on the training epoch and features used. Interestingly, after applying the proposed method, BOLD activation was observed in the frontal lobe region which originally suffered from signal dropout. The validity of this observation needs further verification.

This study proposed a new neural network method for GE-EPI distortion correction based on single subject data acquisition which relies on overfitting onto each individual data. Results show the feasibility of the proposed methods for reducing artifacts in GE-EPI. In addition, BOLD activation is well preserved and potentially can further elucidate activation in artifact dominant regions.

Activation (eg. BOLD task-fMRI) ²

Other Methods

BOLD fMRI ¹

Data analysis

FUNCTIONAL MRI

Machine Learning

MRI

MRI PHYSICS

Workflows