Improved IVIM MRI of Small Lesions in the Liver by Deformable Image Registration with Modality Independent Neighborhood Descriptor
Yihao Guo1, Zhentai Lu1, Yingjie Mei2, Jing Zhang3, Yikai Xu3, Feng Huang4, Ed. X. Wu5,6, and Yanqiu Feng1,5,6

1School of Biomedical Engineering and Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, GuangZhou, China, People's Republic of, 2Philips Healthcare, GuangZhou, China, People's Republic of, 3Department of Medical Imaging Center, Nanfang Hospital, Southern Medical University, GuangZhou, China, People's Republic of, 4Philips Healthcare(Suzhou), Suzhou, China, People's Republic of, 5Laboratory of Biomedical Imaging and Signal Processing, The University of Hong Kong, Hong Kong SAR, China, People's Republic of, 6Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong SAR, China, People's Republic of


Respiration-induced misalignment between multiple b-value liver DW image scan severely reduce the accuracy and stability of IVIM parameter quantification, especially in the presence of small focal lesions. These small lesions usually exhibit significantly differentintensity in different b-value images, but have similar structural information. This work introduces modality independent neighborhood descriptor to extract the structural information of small lesions for improved realignment between multiple b-value images. Preliminary results show that this structure-based registration method can well correct respiration-induced misalignment between multiple b-value images with small lesions, improve the IVIM model fitting quality, and reduce variance in quantified parameters.


The quantitative parameters derived by the intravoxel incoherent motion (IVIM) model from multiple b-value diffusion weighted (DW) images have potential to characterize various liver lesions[1-2]. However, respiration-induced misalignment between multiple b-value liver DW images can severely reduce the accuracy and stability of IVIM parameter quantification, especially in the presence of small focal lesions. Intensity-based deformable image registration methods have been applied to correct motion in liver IVIM MRI [3]. For small lesions, the accurateintensity statistics are difficult to obtain due to limited number of available samples, thus intensity-based mutual information[3] registration method usually cannot achieve accurate realignment. The purpose of this work is to investigate the feasibility of improving the IVIM MRI of small liver lesions using a structure-based registration method based on modality independent neighborhood descriptor (MIND) [4].


In the liver IVIM MRI, small lesions usually have distinct intensities among multiple b-value images but similar structural information. MIND, defined according to local patch similarity, was used to extract the structural information of small lesions. As shown in Figure 1, MIND generated a small matrix for each voxel in the image. It can be observed that the intensity information between two b-value images was significantly different but the structural information represented by MIND was similar between different b-value images for voxels in either the central region or margins of the lesion. After calculating MIND of all voxels, the motion parameters can be found by minimizing the following object function: $$$argmin_T=\sum_xS(I_1(x)),T(I_2(x))^2+\alpha*E_{smooth}(T)$$$ Here, $$$x$$$ denotes voxel index in images, $$$I_1(x)$$$ denotes the reference image, $$$I_2(x)$$$ denotes the float image, $$$ T$$$ denotes the parameters of free-form deformation, $$$S(I(x),J(x))$$$ denotes the similarity measure defined on MIND between images $$$I(x)$$$ and $$$J(x)$$$: $$$S(I(x),J(x))=\frac{1}{|R|}\sum_{r\in R}|MIND(I,x,r)-MIND(J,x,r)|$$$ ,$$$ E_{smooth}(T)$$$ is the smoothness of deformation, $$$|R|$$$ denotes the number of voxels in search window, $$$ r$$$ is the voxel index in $$$ R$$$ and, $$$ α$$$ is a trade-off parameter that balances the smoothness of deformation and the similarity between images. In this work, $$$α$$$ was experientially set to 0.1. Free breathing diffusion datasets of three subjects were acquired on a 3.0T Philips scanner using a single-shot spin-echo echo-planar imaging (EPI) sequence with TR/TE 1600/62ms, matrix 256×256, in-plane resolution 1.46×1.46 mm2, thickness5.0mm, slice gap 0mm, 32 slices and b-values = 0, 10, 30, 60, 100, 150, 400, 600 and 1000 s/mm2. Image of b = 0 was set as the reference image during registration. The IVIM parameters derived before and after registration were compared for method evaluation.


Figure 2 shows the effect of the proposed motion correction method on images with small focal lesions. Before registration, the small lesion appeared on different locations indifferent b-value images. After registration, the lesion in each b-value image was well aligned to that in b = 0 image. Figure 3 plots the mean intensity inside the lesion area and the fitted IVIM model curves against multiple b values. It can be observed that the quantified parameters (D, D* and f) were different before and after registration, and the R-square, which describes the fitting quality, increased from 0.97 to 0.99 after registration. Note that before the registrationthe mean intensity at b = 60 was lower than that at b = 100, which contradicts the fact that signal intensity attenuates with increased b values. Table 1 presents the voxel-wise IVIM fitting results in the focal lesion areas of the three subjects. For subject 1, R-square increased after motion correction with the proposed method. The R-squares for subjects 2 and 3 are already high before motion correction, but slight increase can still be observed after motion correction. The mean values of IVIM model parameters (D, D*, and f) are significantly different with and without motion correction, and the standard deviations of quantified IVIM model parameters consistentlydecreased after registration.

Discussion and Conclusion

The proposed motion correction method employs MIND-described structural information to realign small focal lesions in multiple b-value images of the liver IVIM MRI. The preliminary results on three subjects demonstrate that the proposed method can consistently reduce lesion misalignment, improve fitting quality, and decrease the variance of quantified liver IVIM model parameters. The proposed method has potential to improve the accuracyand reproducibility of free-breathing liver IVIM MRI, especially for small focal lesions. Further evaluation of the proposed method on more subjects is warranted in a future study.


No acknowledgement found.


[1] Zhou,I.Y. MRM 2014; 5:1389-96 [2] Cohen,A.D. MRM 2014; 1:306-11 [3] Mazaheri, Y. Academic radiology 2102; 12:1573-80 [4] Heinrich, M. MIA 2012; 7:1423-35


Fig.1 Illustration of MIND. (a) DW image with b = 0; (b) DW image with b = 1000 s/mm2; (c)and (e) calculated MIND for the red and blue voxels in (a), respectively; (d) and (f) calculated MIND for the red and blue voxels in (b), respectively.

Fig.2 The effectof motion correction on multiple b-value images of subject 1. (a) reference image with b=0. (f) zoom-in view around a small lesion in (a). The blue and red lines represent the lesion edge in b=0 image, and are overlapped on images of different b-values before(b-e) and after(h-k) registration.

Figure 3. The IVIM model fitting result of mean signal in the regions enclosed by red and blue lines in Fig.2. Before registration: R-square = 0.97, D = 2.0×10-3 mm2/s, D* = 31.3×10-3 mm2/s, f = 0.64. After registration: R-square = 0.99, D = 1.5×10-3 mm2/s, D* = 21.7×10-3 mm2/s, f = 0.59.

Table 1 Voxel-wise fitting results (mean±standard deviation) in the lesion regions of three subjects using the IVIM model before and after motion correction.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)