HHS Public Access Author manuscript Author Manuscript

Proc Int Semin Speech Prod. Author manuscript; available in PMC 2016 May 26. Published in final edited form as: Proc Int Semin Speech Prod. 2011 June ; 2011: 17–24.

Deformable Registration of Multimodal Tongue MR Images Jonghye Woo1, Maureen Stone1, and Jerry L. Prince2 1University 2Johns

of Maryland Dental School

Hopkins University, Baltimore, MD

Purpose Author Manuscript

High-resolution magnetic resonance imaging (MRI) and cine MRI of the tongue provide complementary information and play a crucial role in tongue motion analysis. Highresolution MRI provides both boundary and internal details about the anatomy surrounding the vocal tract, but is restricted to a static position. On the other hand, cine MRI can visualize the shape of the vocal tract over time, but internal anatomical details about the tongue and other anatomy is not as clear due to its lower spatial resolution. In this work, we aim at developing fully automated and accurate 3D deformable registration of highresolution MRI with cine MRI. Anatomical and temporal data can be registered to provide correspondences of anatomical tissue points identified in high-resolution MRI, thus potentially allowing for tracking tissue point motion in cine MRI.

Author Manuscript

Method

Author Manuscript

We pre-process both high-resolution and cine MRI images to remove artifacts and create isotropic volumes. The registration algorithm combines affine registration and free-form deformations (FFD) based on uniform cubic B-splines to model the deformable registration by maximizing mutual information (MI) as in [1]. Although MI has been successfully applied to multimodal image registration, it has limitations in that the statistics that are computed from overlap regions may not reflect the local characteristics of salient structural regions (e.g., corners or edges), thereby ignoring spatially meaningful information [2]. To remedy this, a fully automated registration method is presented, utilizing geometric features obtained from scale-space edge localization, and the extracted edges and surrounding areas are used to create an anatomical mask over which to calculate the MI. Additionally, a multiresolution scheme is used to represent coarse-to-fine details of both volumes for fast and robust registration. The energy functional is minimized using a gradient descent method. We used five pairs of 3D axial MRI volumes to validate the algorithm. The image and voxel sizes for high-resolution MRI after pre-processing were 256×256×89 (0.9375×0.9375×0.9375mm3) and for cine MRI, they were 256×256×49 (1.875×1.875×1.875mm3). Registration was performed on both volumes with similar positions: the first time frame of cine MRI that was acquired during speech task of “a geese” and the high-resolution MRI that was acquired at rest. To assess the accuracy of the registration algorithm, we used target registration error (TRE), as defined by Fitzpatrick et al. [3]. For TRE, two expert observers independently selected 3 corresponding anatomical

Woo et al.

Page 2

Author Manuscript

landmarks from each volume including tongue tip, lower lip, and posterior pharynx. We then calculated the mean and standard deviation of pairwise distances of these landmarks after registration.

Results

Author Manuscript

The registration algorithm was performed on an Intel i7 CPU with a clock speed of 1.74 GHz and 8 GByte memory. The mean computation time for the whole process was 10±4 min. Figure 1 shows one example result before and after registration. Table 1 lists the mean and standard deviation of TRE and inter-observer variability. The TRE results show that the proposed method provided accurate results compared to the conventional method using pure FFD method. Of note, selecting anatomical landmarks is a difficult task and is of great importance in assessing the accuracy of registration algorithm. This is because quantitative evaluation of the registration is challenging as there is no true gold standard other than visual judgment, which is also associated with inter-observer variability. In the present study, the inter-observer variability was high especially in the lower lip but the TRE obtained using the proposed method was comparable to inter-observer variability (p=NS).

Conclusion Fully automated 3D deformable registration of high-resolution MRI and cine MRI of tongue can be performed accurately with average error of TRE comparable to inter-observer variability. The proposed approach allows accurate tissue point localization or muscle and structural definition in cine MRI. Additionally, this method could enhance the interpretation of speech gestures and disorders.

Author Manuscript

References 1. Rueckert D, et al. Nonrigid registration using free-form deformations: application to breast MR images. IEEE Trans Med Imaging. 1999; 18(8):712–21. [PubMed: 10534053] 2. Pluim JP, Maintz JB, Viergever MA. Image registration by maximization of combined mutual information and gradient information. IEEE Trans Med Imaging. 2000; 19(8):809–14. [PubMed: 11055805] 3. Fitzpatrick JM, West JB, Maurer CR Jr. Predicting error in rigid-body point-based registration. IEEE Trans Med Imaging. 1998; 17(5):694–702. [PubMed: 9874293]

Author Manuscript Proc Int Semin Speech Prod. Author manuscript; available in PMC 2016 May 26.

Woo et al.

Page 3

Author Manuscript

Figure 1.

Author Manuscript

Cine MRI (a) was registered to high-resolution MRI (b). Original data misalignment was large as depicted in (c). Fusion result using deformable registration (e) showed slightly better alignment compared to that using affine registration (d). Deformation grid combining affine and deformable registration is shown in (f). The green arrows show the misalignment of affine registration

Author Manuscript Author Manuscript Proc Int Semin Speech Prod. Author manuscript; available in PMC 2016 May 26.

Author Manuscript Affine registration 8.3±2.7 5.7±2.7 3.0±2.0 5.7±3.2

Before registration 11.6±7.0 7.6±4.6 3.4±1.7 7.5±5.7

Tongue tip

Lower lip

Posterior pharynx

Average

4.7±2.1

3.3±1.5

5.5±2.0

5.4±2.2

Conventional method

Author Manuscript

TRE (mm)

3.9±2.2

3.0±1.8

4.8±2.9

3.8±2.0

Proposed method

3.7±2.8

3.0±2.1

5.3±3.3

2.8±2.5

Observer variability (mm)

Author Manuscript

TRE and observer variability

Author Manuscript

Table 1 Woo et al. Page 4

Proc Int Semin Speech Prod. Author manuscript; available in PMC 2016 May 26.

Deformable Registration of Multimodal Tongue MR Images.

Deformable Registration of Multimodal Tongue MR Images. - PDF Download Free
91KB Sizes 0 Downloads 14 Views