Elsevier

Medical Image Analysis

Volume 39, July 2017, Pages 124-132
Medical Image Analysis

Motion-robust parameter estimation in abdominal diffusion-weighted MRI by simultaneous image registration and model estimation

https://doi.org/10.1016/j.media.2017.04.006Get rights and content

Highlights

  • Motion-robust parameter estimation is needed in abdominal diffusion-weighted MRI.

  • We propose a novel simultaneous image registration and model estimation (SIR-ME) framework.

  • SIR-ME utilizes dependency of acquired b-value images along the diffusion-weighting dimension.

  • Jointly estimates transformations for the non-rigid alignment and signal decay model parameters.

  • Increases precision of parameters, improves discrimination of normal and diseased bowel.

Abstract

Quantitative body DW-MRI can detect abdominal abnormalities as well as monitor response-to-therapy for applications including cancer and inflammatory bowel disease with increased accuracy. Parameter estimates are obtained by fitting a forward model of DW-MRI signal decay to the observed data acquired with several b-values. The DW-MRI signal decay models typically used do not account for respiratory, cardiac and peristaltic motion, however, which may deteriorate the accuracy and robustness of parameter estimates. In this work, we introduce a new model of DW-MRI signal decay that explicitly accounts for motion. Specifically, we estimated motion-compensated model parameters by simultaneously solving image registration and model estimation (SIR-ME) problems utilizing the interdependence of acquired volumes along the diffusion-weighting dimension. To accomplish this, we applied the SIR-ME model to the in-vivo DW-MRI data sets of 26 Crohn’s disease (CD) patients and achieved improved precision of the estimated parameters by reducing the coefficient of variation by 8%, 24% and 8% for slow diffusion (D), fast diffusion (D*) and fast diffusion fraction (f) parameters respectively, compared to parameters estimated with independent registration in normal-appearing bowel regions. Moreover, the parameters estimated with the SIR-ME model reduced the error rate in classifying normal and abnormal bowel loops to 12% for D and 10% for f parameter with a reduction in error rate by 13% and 11% for D and f parameters, respectively, compared to the error rate in classifying parameter estimates obtained with independent registration. The experiments in DW-MRI of liver in 20 subjects also showed that the SIR-ME model improved the precision of parameter estimation by reducing the coefficient of variation to 7% for D, 23% for D*, and 8% for the f parameter. Using the SIR-ME model, the coefficient of variation was reduced by 4%, 14% and 6% for D, D* and f parameters, respectively, compared to parameters estimated with independent registration. These results demonstrate that the proposed SIR-ME model improves the accuracy and robustness of quantitative body DW-MRI in characterizing tissue microstructure.

Introduction

Quantitative diffusion-weighted MRI (DW-MRI) parameters have been increasingly used for the characterization of abnormalities in tissue microstructure of liver, spleen and bowel (Chavhan, AlSabban, Babyn, 2014, Jang, Kim, Hwang, Lee, Kang, Lee, Choi, 2014, Yoon, Lee, Yu, Kiefer, Han, Choi, 2014, Oto, Kayhan, Williams, Fan, Yun, Arkani, Rubin, 2011). The water molecule mobility attenuates the diffusion-weighted MR signal according to the b-value used in the acquisition. Typically, DW-MRI images are acquired at multiple b-values. A signal decay model is then fitted to the measured signal and the signal decay rate parameters are computed.

The mobility of water molecules in tissue microenvironments changes in the presence of abnormal tissue because of modified cellularity, cell membrane integrity and micro-capillary perfusion. Changes in tissue microenvironments can be identified quantitatively using signal decay model parameters. For instance, malignant lesions are expected to exhibit restricted diffusion, with a lower decay rate of diffusion due to reduced extracellular space; while benign lesions generally have more extracellular space that allows for more normal diffusion, with a higher decay rate of diffusion. However, reproducible and precise parameter estimation techniques are required to increase the sensitivity and specificity of these parameters to detect abnormalities and monitor response-to-therapy, which, in turn, is expected to increase the utility of quantitative DW-MRI in clinical care settings.

Several models have been proposed to quantify the signal decay in the DW-MRI images. A mono-exponential signal decay model (Koh and Collins, 2007), which encapsulates the multiple signal decay rates by a single parameter called the “apparent diffusion coefficient (ADC)”. ADC model is used most often due to its robustness and ease of image acquisition and parameter computation. However, this simplified model precludes the independent characterization of multiple diffusion scales– a process essential to accurately quantifying the biological phenomena taking place inside the tissue of interest. The bi-exponential signal decay model considers both slow and fast components of signal decay according to the intra-voxel incoherent motion (IVIM) theory (Le Bihan et al., 1988). The signal decay at high b-values, which is associated with the slow-diffusion, reflects the mobility of water molecules in the tissue. The signal decay at low b-values, associated with the fast-diffusion component, is an indicator of micro-capillary perfusion. The IVIM model has 3 parameters: a slow diffusion coefficient (D), a fast diffusion coefficient (D*), and a fraction coefficient (f) reflecting the proportion of fast diffusion spins at b = 0 s/mm2.

Both slow and fast diffusion in biological tissue are heterogeneous processes that occur over a broad range of time scales due to widely varying cell structures, vessel sizes and flow rates. Recently, a more accurate probabilistic model of diffusion has been proposed that considers a full characterization of the distribution of diffusion scales that attenuate the DW-MRI signal using a two-component probability mixture model. A spatial homogeneity prior has been added to this model (Kurugol et al., 2014) to obtain reliable estimates of parameters using the Fusion Bootstrap Moves (FBM) algorithm proposed by Freiman et al. (2013) for the spatially-constrained IVIM model. This spatially-constrained probability model of incoherent motion (SPIM) has been shown to characterize the entire scale of diffusion reflecting the tissue microstructure while increasing the precision and reproducibility of parameter estimation in low signal-to-noise ratio (SNR) DW-MRI images (Kurugol et al., 2016). None of these signal decay models, however, consider the presence of the respiratory, cardiac and peristalsis motion, which causes misalignment between image volumes, acquired at multiple b-values and deteriorates the accuracy and robustness of parameter estimation.

Previous techniques for motion compensation include breath-holding, gating, and respiratory or cardiac triggering. These techniques have disadvantages, however, such as increased scan time and a need for the patient’s cooperation. Furthermore, none of these approaches entirely correct for motion. While breath-holding methods can be used to improve the robustness of DW-MRI data (Kandpal, Sharma, Madhusudhan, Kapoor, 2009, Kwee, Takahara, Koh, Nievelstein, Luijten, 2008), only a limited number of b-values can be obtained within a breath-hold, and this requires the patient’s cooperation. Naturally, such a method is not suitable for young children who cannot hold their breath, or who must be imaged under sedation. By contrast, free-breathing DW-MRI has the effect of signal-averaging over large regions of interest, resulting in accurate parameter estimations for large homogenous regions, which, in turn, may improve SNR by using multiple signal acquisitions. However, this technique results in reduced accuracy for small and heterogeneous lesions (Koh, Collins, 2007, Koh, Collins, Orton, 2011) and scan time increases linearly with number of excitations. Respiratory triggering methods have also been shown to reduce motion artifact, but at the expense of increased scanning times (Taouli, Sandberg, Stemmer, Parikh, Wong, Xu, Lee, 2009, Taouli, Koh, 2010). In addition, the triggering technique does not always perform well if the respiratory rhythm is irregular as in the case of anxious awake children who are breathing rapidly or irregularly. Residual motion artifacts still remain in triggered respiratory scans and may, as a consequence, decrease the precision of diffusion parameters. Especially, the estimation of the micro-capillary perfusion contribution (f) demonstrated a relatively large variability (Eisenberger et al., 2013), which may be due, in part, to residual motion effects.

Another alternative is post acquisition motion compensation based on image registration, to bring the volumes acquired at different b-values into the same physical coordinate space before fitting a signal decay model (Guyader, Bernardin, Douglas, Poot, Niessen, Klein, 2014, Mazaheri, Do, Shukla-Dave, Deasy, Lu, Akin, 2012). However, each b-value image has different contrast; as a result, independent registration of different b-value images to a b = 0 s/mm2 image may not be very accurate, especially for high b-value images where the signal is significantly attenuated and the signal to noise ratio is low.

Several physiological model driven registration methods were used for motion compensation in dynamic contrast-enhanced MR imaging (Buonaccorsi, O’Connor, Caunce, Roberts, Cheung, Watson, Davies, Hope, Jackson, Jayson, et al., 2007, Bhushan, Schnabel, Risser, Heinrich, Brady, Jenkinson, 2011). Recently, a different data driven method was proposed by Huizinga et al. (2016), which registers quantitative MR images without using any predefined model by utilizing a PCA-based groupwise image registration technique. However, the PCA-based representation is only applicable to data from a simplified single exponential decay rather than data with an underlying complex signal decay composed of a bi-modal distribution of fast and slow diffusion components.

In this work, we introduce a simultaneous image registration and model estimation (SIR-ME) framework for motion-compensated parameter estimation of both fast and slow diffusion components in DW-MRI. This paper extends our model-driven motion compensation and parameter estimation framework for abdominal diffusion imaging previously presented at the MICCAI 2015 conference (Kurugol et al., 2015) by providing a more detailed description of the model based on additional experiments on a larger dataset that more extensively analyze the precision of the parameter estimation using the proposed approach in comparison to previous methods in normal and diseased bowel regions of Crohn’s disease patients. We also included additional experiments on DW-MRI dataset of upper abdomen to test the performance of the proposed approach in liver regions of 20 subjects.

Our primary contribution consists of a new signal decay model that characterizes the entire scale of diffusion robustly while considering the likelihood of patient motion and compensating for it, with the goal of obtaining robust parameter estimations. The SIR-ME framework described in this paper utilizes the dependency of acquired volumes in the diffusion-weighting dimension, where images are related to each other through the signal decay model. Incorporating this information as additional prior information into the motion compensation framework is expected to improve the performance of registration for motion-compensation. The SIR-ME solver utilizes this information and jointly estimates transformations for the non-rigid alignment of images; reconstructs high SNR registered diffusion images; and estimates signal decay model parameters. This novel, joint parameter estimation method is solved iteratively to obtain improved parameter estimation. In parallel, it corrects for the effects of motion and reconstructs motion-compensated image volumes.

Section snippets

Intra-voxel incoherent motion model (IVIM)

In quantitative DW-MRI, images are acquired at multiple (i=1..N) b-values. In the absence of motion, a signal decay model is then fitted to the measured signal at multiple b-values (bi), and the model parameters are estimated. The intra-voxel incoherent motion model of DW-MRI signal decay proposed by Le Bihan and Turner (1991) assumes a bi-exponential signal decay function to model both inherent slow diffusion due to Brownian motion of water molecules, and fast diffusion due to bulk motion of

Experiments

We have tested the performance of the proposed motion-compensated, model estimation framework on in-vivo DW-MRI data of lower abdomen in 26 Crohn’s disease (CD) patients and in-vivo DW-MRI of upper abdomen in 20 patients using a 1.5T MRI scanner (Magnetom Avanto, Siemens Medical Solutions, Erlangen, Germany) with an 8-channel, body matrix receive coil. Free-breathing, single-shot echo-planar imaging was performed using the following parameters: repetition time/echo time (TR/TE)= 7500/77 ms;

Results

First row of Fig. 1 shows 3 different b-value images indicating the labeled bowel wall with Crohn’s disease and the normal looking bowel wall. The second and third row of Fig. 1 compares resultant parameter maps of D and f parameters estimated using w/o registration, with registration, and SIR-ME methods. SIR-ME model results in improved parameter maps with less noise due to the effect of motion compensation. The left panel of the last row in Fig. 1 shows an image column selected around the CD

Conclusions and discussion

Quantitative abdominal DW-MRI is increasingly used in evaluating patients with a known or suspected disease in liver, spleen and bowel. Specific applications include distinguishing between diseased and normal-appearing regions (e.g., areas of inflamed bowel characteristic of Crohn’s disease (Neubauer et al., 2013)) and monitoring response-to-therapy for various cancers of the abdomen. Our ability to successfully utilize quantitative DW-MRI in routine clinical practice depends, however, on

References (29)

  • U. Eisenberger et al.

    Living renal allograft transplantation: diffusion-weighted MR imaging in longitudinal follow-up of the donated and the remaining kidney

    Radiology

    (2013)
  • M. Freiman et al.

    Quantitative body DW-MRI biomarkers uncertainty estimation using unscented wild-bootstrap

    Medical Image Computing and Computer-Assisted Intervention–MICCAI 2011

    (2011)
  • J.-M. Guyader et al.

    Influence of image registration on apparent diffusion coefficient images computed from free-breathing diffusion MRI of the abdomen

    JMRI

    (2014)
  • K.M. Jang et al.

    Differentiation of malignant from benign focal splenic lesions: added value of diffusion-weighted MRI

    Am. J. Roentgenol.

    (2014)
  • Cited by (18)

    • An efficient two-step multi-organ registration on abdominal CT via deep-learning based segmentation

      2021, Biomedical Signal Processing and Control
      Citation Excerpt :

      Freiman et al. [33] employed the non-rigid Thirion’s demons with a local affine regularization term to obtain an adaptive smoothing deformation field for abdominal CT images registration. Kurugol et al. [34] combined a cross-correlation similarity measure with an error norm to form local block-matching based transformation for abdominal diffusion-weighted magnetic resonance imaging (MRI) registration. In this paper, a two-step non-rigid registration strategy is presented for 3D abdominal CT images.

    • Modeling dynamic radial contrast enhanced MRI with linear time invariant systems for motion correction in quantitative assessment of kidney function

      2021, Medical Image Analysis
      Citation Excerpt :

      These reduce the volume-to-volume variability of the registration, but might not capture the sudden changes in position of the subject. An alternative approach to balance stability and rapid motion consists of registering individual volumes using the tracer kinetic (TK) model Adluru et al. (2006); Buonaccorsi et al. (2005, 2007) or the signal decay model Kurugol et al. (2017b) itself. This approach allows to compensate for rapid bulk motion —since each volume is registered independently— while introducing robustness to artifacts with the noise free templates generated with the signal model.

    • A review on 3D deformable image registration and its application in dose warping

      2020, Radiation Medicine and Protection
      Citation Excerpt :

      Conventional DIR methods have been applied in many clinical tasks, and all of them could be potential applications for DL-based DIR methods. For example, Kurugol et al99 used DIR in abdominal diffusion-weighted MRI reconstruction by registering images with different b values; Wang et al100 did segmentation with DIR by deforming a segmented atlas to other subjects; Ebner et al101 used DIR in abdominal image super-resolution by correcting organ motions. In principle, DL-based DIR models could also be used in abovementioned applications.

    • Robust estimation of the apparent diffusion coefficient invariant to acquisition noise and physiological motion

      2018, Magnetic Resonance Imaging
      Citation Excerpt :

      On the other hand, groupwise approaches are based on an image reference built out of the whole image set, so the bias mentioned above vanishes. Recently, spatially-constrained approaches have also been proposed for liver DWI non-rigid registration (see Refs. [14, 15]) grounded on Markov Random Field (MRF), showing an increase in robustness of diffusion parameter estimation by enforcing boundary consistency. Groupwise approaches that use a MRF have also been described [16].

    • qDWI-Morph: Motion-Compensated Quantitative Diffusion-Weighted MRI Analysis for Fetal Lung Maturity Assessment

      2023, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    View all citing articles on Scopus

    This work is supported by Crohn’s and Colitis Foundation of America’s Career Development Award, the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) of the NIH under award R01DK100404 and by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) of the NIH under award R01EB019483.

    View full text