Quantitative susceptibility mapping using deep neural network: QSMnet

Yoon, Jaeyeon; Gong, Enhao; Chatnuntawech, Itthi; Bilgic̦, Berkin; Lee, Jingu; Jung, Woojin; Ko, Jingyu; Jung, Hosan; Setsompop, Kawin; Zaharchuk, Greg; Kim, Eung Yeop; Pauly, John M.; Lee, Jong-Ho

doi:10.1016/j.neuroimage.2018.06.030

Cited by 150 publications

(184 citation statements)

References 45 publications

(45 reference statements)

Supporting

Mentioning

160

Contrasting

Order By: Relevance

“…Recently, deep learning (DL) using a neural network has shown remarkable potential for similar problems in which model‐based analytical approaches are difficult to apply . The method can learn a nonlinear mapping from an input space to an output space when enough dataset pairs are given.…”

mentioning

confidence: 99%

Data‐driven synthetic MRI FLAIR artifact correction via deep neural network

Ryu

Nam

Gho³

et al. 2019

Magnetic Resonance Imaging

View full text Add to dashboard Cite

Background FLAIR (fluid attenuated inversion recovery) imaging via synthetic MRI methods leads to artifacts in the brain, which can cause diagnostic limitations. The main sources of the artifacts are attributed to the partial volume effect and flow, which are difficult to correct by analytical modeling. In this study, a deep learning (DL)‐based synthetic FLAIR method was developed, which does not require analytical modeling of the signal. Purpose To correct artifacts in synthetic FLAIR using a DL method. Study Type Retrospective. Subjects A total of 80 subjects with clinical indications (60.6 ± 16.7 years, 38 males, 42 females) were divided into three groups: a training set (56 subjects, 62.1 ± 14.8 years, 25 males, 31 females), a validation set (1 subject, 62 years, male), and the testing set (23 subjects, 57.3 ± 20.4 years, 13 males, 10 females). Field Strength/Sequence 3 T MRI using a multiple‐dynamic multiple‐echo acquisition (MDME) sequence for synthetic MRI and a conventional FLAIR sequence. Assessment Normalized root mean square (NRMSE) and structural similarity (SSIM) were computed for uncorrected synthetic FLAIR and DL‐corrected FLAIR. In addition, three neuroradiologists scored the three FLAIR datasets blindly, evaluating image quality and artifacts for sulci/periventricular and intraventricular/cistern space regions. Statistical Tests Pairwise Student's t‐tests and a Wilcoxon test were performed. Results For quantitative assessment, NRMSE improved from 4.2% to 2.9% (P < 0.0001) and SSIM improved from 0.85 to 0.93 (P < 0.0001). Additionally, NRMSE values significantly improved from 1.58% to 1.26% (P < 0.001), 3.1% to 1.5% (P < 0.0001), and 2.7% to 1.4% (P < 0.0001) in white matter, gray matter, and cerebral spinal fluid (CSF) regions, respectively, when using DL‐corrected FLAIR. For qualitative assessment, DL correction achieved improved overall quality, fewer artifacts in sulci and periventricular regions, and in intraventricular and cistern space regions. Data Conclusion The DL approach provides a promising method to correct artifacts in synthetic FLAIR. Level of Evidence: 4 Technical Efficacy: Stage 1 J. Magn. Reson. Imaging 2019;50:1413–1423.

show abstract

mentioning

confidence: 99%

Data‐driven synthetic MRI FLAIR artifact correction via deep neural network

Ryu

Nam

Gho³

et al. 2019

Magnetic Resonance Imaging

View full text Add to dashboard Cite

show abstract

“…This deep neural network was trained with local field maps and the corresponding COSMOS results. Compared with results from conventional QSM reconstruction algorithms, QSMnet showed faster reconstruction speed (40 times faster than MEDI) and higher consistency (lower residual error than TKD and MEDI) …”

Section: Introductionmentioning

confidence: 86%

“…To conduct deep learning‐based image to image operation, QSMnet, which is a modified version of U‐net structure with 3D inputs and outputs, was utilized. The detailed network architecture used in our study is illustrated in Fig.…”

Section: Methodsmentioning

confidence: 99%

“…The network was modified to exclude the innermost convolution, deconvolution and feature concatenation layers with 512 channels. For QSMnet training, the same loss function proposed by Yoon et al was used . The total loss function was defined as follows,

\begin{matrix} Loss = & {normalw}_{1} \cdot (m o d e l l o s s) + w_{2} \cdot (L 1 l o s s) + w_{3} \cdot (g r a d i e n t l o s s) \\ = & {normalw}_{1} {(∥∥ d * χ - d * y)}_{1} + w_{2} {(∥∥ χ - y)}_{1} \\ + w_{3} ({∥(||, \nabla, (normald * normalχ)) - (||, \nabla, (normald * normaly))∥}_{1} + {∥(||, \nabla, χ) - (||, \nabla, y)∥}_{1}) \\ {normalw}_{1}, {normalw}_{2} {and w}_{3} : Weighting factors, d:dipole kernel, \\ χ : output of QSMnet, y: output label \end{matrix}

…”

Section: Methodsmentioning

confidence: 99%

“…To reduce blocky artifacts that commonly occurred when using a nonoverlapping patch method the patch was generated with an overlapping scheme of 66% overlap between adjacent patches. As proposed in the previous study a whole 3D local field map, not patches, was applied to the input of the QSMnet for test data or for inference …”

Section: Methodsmentioning

confidence: 99%

See 2 more Smart Citations

A preliminary attempt to visualize nigrosome 1 in the substantia nigra for Parkinson’s disease at 3T: An efficient susceptibility map‐weighted imaging (SMWI) with quantitative susceptibility mapping using deep neural network (QSMnet)

2020

Medical Physics

View full text Add to dashboard Cite

Purpose Visibility of nigrosome 1 in the substantia nigra (SN) is used as an MR imaging biomarker for Parkinson's disease. Because of lower susceptibility induced tissue contrast and SNR visualization of the SN pars compacta (SNPC) using conventional imaging technique in the clinical field strength (≤3T) has been limited. Susceptibility map‐weighted imaging (SMWI) has been proposed to visualize SNPC at 3T. To better visualize nigrosome 1 and SN areas using SMWI, accurate estimation of the quantitative susceptibility mapping (QSM) map is essential. In SMWI processing, however, QSM processing time using conventional algorithms is the most time‐consuming step and may limit clinical use. In this study, we introduce an efficient SMWI processing approach using the deep neural network (QSMnet). To improve the processing speed of SMWI while maintaining similar image quality to that obtained with the conventional method, QSMnet was applied to generate a susceptibility mask for SMWI processing. Methods To conduct a deep learning‐based image to image operation modified QSMnet was utilized. The network was trained with in vivo MR data from 57 healthy controls. To validate SMWI results from QSMnet, four datasets from healthy controls were used as the test datasets. As a preliminary attempt to explore the clinical applicability, Parkinson’s disease patient data were additionally tested. The SWMI images generated by QSMnet and conventional model‐based QSM outputs were compared. To validate SMWI results, region of interest (ROI) analysis was performed. The mean signal values at the nigrosome 1 and the surrounding regions were measured to calculate contrast‐to‐noise ratio (CNR). Results The experimental results confirmed that the proposed approach using QSMnet provided similar QSM and SMWI compared to that obtained with conventional iLSQR while the processing speed was much improved (5.4 times faster). The QSM results from QSMnet show similar tissue contrast with results from iLSQR. When compared, the absolute mean intensity difference between two methods near the nigrosome 1 was 0.015 ppm. SMWI results using susceptibility masks from QSMnet demonstrated signal distribution and tissue contrast that was comparable with those results from the conventional iLSQR method. The absolute difference maps of SMWI were calculated to show the similarity between the two methods. The overall mean absolute difference value in the presented ROIs obtained from healthy controls (n = 4) and a PD patient (n = 1) were 2.31 and 1.81, respectively. Mean CNR values (10 ROIs; n = 5; including both sides; 1.42 for QSMnet; 1.43 for iLSQR) between SN and nigrosome 1 in SMWI results obtained by ROI analysis were similar (P = 0.724). Conclusions In this study, we assessed an efficient approach for SMWI visualizing SN and nigrosome 1 on 3T. QSMnet provides a similar SMWI image to that obtained with the conventional iterative QSM algorithm and improves QSM processing speed by avoiding iterative computation. Since QSM is the most time‐consuming step of SMWI proc...

show abstract

Quantitative Susceptibility Mapping: Technical Considerations and Clinical Applications in Neuroimaging

Vinayagamani¹,

Sheelakumari²,

Sabarish³

et al. 2020

Magnetic Resonance Imaging

View full text Add to dashboard Cite

Quantitative susceptibility mapping (QSM) is a novel magnetic resonance imaging (MRI) technique for quantifying the spatial distribution of magnetic susceptibility within an object or tissue. Recently, QSM has been widely used to study various dominant magnetic susceptibility sources in the brain, including iron and calcium. In addition, the method enables mapping of the cerebral metabolic rate of oxygen, which could act as a new metabolic biomarker for diseases that involve disruption of the brain's oxygen supply. Thus, the clinical applications of QSM are wide-reaching and hold great promise as imaging biomarkers for studying several neurological diseases. This review aims to summarize the physical concepts and potential clinical applications of QSM in neuroimaging. Evidence Level: 5 Technical Efficacy: Stage 2

show abstract

Quantitative susceptibility mapping using deep neural network: QSMnet

Cited by 150 publications

References 45 publications

Data‐driven synthetic MRI FLAIR artifact correction via deep neural network

Data‐driven synthetic MRI FLAIR artifact correction via deep neural network

A preliminary attempt to visualize nigrosome 1 in the substantia nigra for Parkinson’s disease at 3T: An efficient susceptibility map‐weighted imaging (SMWI) with quantitative susceptibility mapping using deep neural network (QSMnet)

Quantitative Susceptibility Mapping: Technical Considerations and Clinical Applications in Neuroimaging

Contact Info

Product

Resources

About