Dataset Information

Deep neural network for water/fat separation: Supervised training, unsupervised training, and no training.

ABSTRACT:

Purpose

To use a deep neural network (DNN) for solving the optimization problem of water/fat separation and to compare supervised and unsupervised training.

Methods

The current T2∗ -IDEAL algorithm for solving water/fat separation is dependent on initialization. Recently, DNN has been proposed to solve water/fat separation without the need for suitable initialization. However, this approach requires supervised training of DNN using the reference water/fat separation images. Here we propose 2 novel DNN water/fat separation methods: 1) unsupervised training of DNN (UTD) using the physical forward problem as the cost function during training, and 2) no training of DNN using physical cost and backpropagation to directly reconstruct a single dataset. The supervised training of DNN, unsupervised training of DNN, and no training of DNN methods were compared with the reference T2∗ -IDEAL.

Results

All DNN methods generated consistent water/fat separation results that agreed well with T2∗ -IDEAL under proper initialization.

Conclusion

The water/fat separation problem can be solved using unsupervised deep neural networks.

SUBMITTER: Jafari R

PROVIDER: S-EPMC7809709 | biostudies-literature | 2021 Apr

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Deep neural network for water/fat separation: Supervised training, unsupervised training, and no training.

Jafari Ramin R Spincemaille Pascal P Zhang Jinwei J Nguyen Thanh D TD Luo Xianfu X Cho Junghun J Margolis Daniel D Prince Martin R MR Wang Yi Y

Magnetic resonance in medicine 20201026 4

<h4>Purpose</h4>To use a deep neural network (DNN) for solving the optimization problem of water/fat separation and to compare supervised and unsupervised training.<h4>Methods</h4>The current T2∗ -IDEAL algorithm for solving water/fat separation is dependent on initialization. Recently, DNN has been proposed to solve water/fat separation without the need for suitable initialization. However, this approach requires supervised training of DNN using the reference water/fat separation images. Here w ...[more]

PMID: 33107127

Similar Datasets

Project description:Chemical-shift encoded MRI (CSE-MRI) is a widely used technique for the study of body composition and metabolic disorders, where derived fat and water signals enable the quantification of adipose tissue and muscle. The UK Biobank is acquiring whole-body Dixon MRI (a specific implementation of CSE-MRI) for over 100,000 participants. Current processing methods associated with large whole-body volumes are time intensive and prone to artifacts during fat-water separation performed by the scanner, making quantitative analysis challenging. The most common artifacts are fat-water swaps, where the labels are inverted at the voxel level. It is common for researchers to discard swapped data (generally around 10%), which is wasteful and may lead to unintended biases. Given the large number of whole-body Dixon MRI acquisitions in the UK Biobank, thousands of swaps are expected to be present in the fat and water volumes from image reconstruction performed on the scanner. If they go undetected, errors will propagate into processes such as organ segmentation, and dilute the results in population-based analyses. There is a clear need for a robust method to accurately separate fat and water volumes in big data collections like the UK Biobank. We formulate fat-water separation as a style transfer problem, where swap-free fat and water volumes are predicted from the acquired Dixon MRI data using a conditional generative adversarial network, and introduce a new loss function for the generator model. Our method is able to predict highly accurate fat and water volumes free from artifacts in the UK Biobank. We show that our model separates fat and water volumes using either single input (in-phase only) or dual input (in-phase and opposed-phase) data, with the latter producing superior results. Our proposed method enables faster and more accurate downstream analysis of body composition from Dixon MRI in population studies by eliminating the need for visual inspection or discarding data due to fat-water swaps.Supplementary informationThe online version contains supplementary material available at 10.1186/s40537-022-00677-1.

Project description:ObjectiveTo accurately estimate liver PDFF from chemical shift-encoded (CSE) MRI using a deep learning (DL)-based Multi-Decoder Water-Fat separation Network (MDWF-Net), that operates over complex-valued CSE-MR images with only 3 echoes.MethodsThe proposed MDWF-Net and a U-Net model were independently trained using the first 3 echoes of MRI data from 134 subjects, acquired with conventional 6-echoes abdomen protocol at 1.5 T. Resulting models were then evaluated using unseen CSE-MR images obtained from 14 subjects that were acquired with a 3-echoes CSE-MR pulse sequence with a shorter duration compared to the standard protocol. Resulting PDFF maps were qualitatively assessed by two radiologists, and quantitatively assessed at two corresponding liver ROIs, using Bland Altman and regression analysis for mean values, and ANOVA testing for standard deviation (STD) (significance level: .05). A 6-echo graph cut was considered ground truth.ResultsAssessment of radiologists demonstrated that, unlike U-Net, MDWF-Net had a similar quality to the ground truth, despite it considered half of the information. Regarding PDFF mean values at ROIs, MDWF-Net showed a better agreement with ground truth (regression slope = 0.94, R2 = 0.97) than U-Net (regression slope = 0.86, R2 = 0.93). Moreover, ANOVA post hoc analysis of STDs showed a statistical difference between graph cuts and U-Net (p < .05), unlike MDWF-Net (p = .53).ConclusionMDWF-Net showed a liver PDFF accuracy comparable to the reference graph cut method, using only 3 echoes and thus allowing a reduction in the acquisition times.Clinical relevance statementWe have prospectively validated that the use of a multi-decoder convolutional neural network to estimate liver proton density fat fraction allows a significant reduction in MR scan time by reducing the number of echoes required by 50%.Key points• Novel water-fat separation neural network allows for liver PDFF estimation by using multi-echo MR images with a reduced number of echoes. • Prospective single-center validation demonstrated that echo reduction leads to a significant shortening of the scan time, compared to standard 6-echo acquisition. • Qualitative and quantitative performance of the proposed method showed no significant differences in PDFF estimation with respect to the reference technique.

Project description:Magnetic susceptibility source separation (χ-separation), an advanced quantitative susceptibility mapping (QSM) method, enables the separate estimation of paramagnetic and diamagnetic susceptibility source distributions in the brain. Similar to QSM, it requires solving the ill-conditioned problem of dipole inversion, suffering from so-called streaking artifacts. Additionally, the method utilizes reversible transverse relaxation ( R2'=R2*-R2$$ {R}_2^{\prime }={R}_2^{\ast }-{R}_2 $$ ) to complement frequency shift information for estimating susceptibility source concentrations, requiring time-consuming data acquisition for R2$$ {R}_2 $$ (e.g., multi-echo spin-echo) in addition to multi-echo GRE data for R2*$$ {R}_2^{\ast } $$ . To address these challenges, we develop a new deep learning network, χ-sepnet, and propose two deep learning-based susceptibility source separation pipelines, χ-sepnet- R2'$$ {R}_2^{\prime } $$ for inputs with multi-echo GRE and multi-echo spin-echo (or turbo spin-echo) and χ-sepnet- R2*$$ {R}_2^{\ast } $$ for input with multi-echo GRE only. The neural network is trained using multiple head orientation data that provide streaking artifact-free labels, generating high-quality χ-separation maps. The evaluation of the pipelines encompasses both qualitative and quantitative assessments in healthy subjects, and visual inspection of lesion characteristics in multiple sclerosis patients. The susceptibility source-separated maps of the proposed pipelines delineate detailed brain structures with substantially reduced artifacts compared to those from the conventional regularization-based reconstruction methods. In quantitative analysis, χ-sepnet- R2'$$ {R}_2^{\prime } $$ achieves the best outcomes followed by χ-sepnet- R2*$$ {R}_2^{\ast } $$ , outperforming the conventional methods. When the lesions of multiple sclerosis patients are classified into subtypes, most lesions are identified as the same subtype in the maps from χ-sepnet- R2'$$ {R}_2^{\prime } $$ and χ-sepnet- R2*$$ {R}_2^{\ast } $$ (paramagnetic susceptibility: 99.6% and diamagnetic susceptibility: 98.4%; both out of 250 lesions). The χ-sepnet- R2*$$ {R}_2^{\ast } $$ pipeline, which only requires multi-echo GRE data, has demonstrated its potential to offer broad clinical and scientific applications, although further evaluations for various diseases and pathological conditions are necessary.

Project description:Inferior temporal (IT) cortex in human and nonhuman primates serves visual object recognition. Computational object-vision models, although continually improving, do not yet reach human performance. It is unclear to what extent the internal representations of computational models can explain the IT representation. Here we investigate a wide range of computational model representations (37 in total), testing their categorization performance and their ability to account for the IT representational geometry. The models include well-known neuroscientific object-recognition models (e.g. HMAX, VisNet) along with several models from computer vision (e.g. SIFT, GIST, self-similarity features, and a deep convolutional neural network). We compared the representational dissimilarity matrices (RDMs) of the model representations with the RDMs obtained from human IT (measured with fMRI) and monkey IT (measured with cell recording) for the same set of stimuli (not used in training the models). Better performing models were more similar to IT in that they showed greater clustering of representational patterns by category. In addition, better performing models also more strongly resembled IT in terms of their within-category representational dissimilarities. Representational geometries were significantly correlated between IT and many of the models. However, the categorical clustering observed in IT was largely unexplained by the unsupervised models. The deep convolutional network, which was trained by supervision with over a million category-labeled images, reached the highest categorization performance and also best explained IT, although it did not fully explain the IT data. Combining the features of this model with appropriate weights and adding linear combinations that maximize the margin between animate and inanimate objects and between faces and other objects yielded a representation that fully explained our IT data. Overall, our results suggest that explaining IT requires computational features trained through supervised learning to emphasize the behaviorally important categorical divisions prominently reflected in IT.

Dataset Information

Deep neural network for water/fat separation: Supervised training, unsupervised training, and no training.

Purpose

Methods

Results

Conclusion

Publications

Deep neural network for water/fat separation: Supervised training, unsupervised training, and no training.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets