Reconstruction of fetal brain MRI with intensity matching and complete outlier removal.
Kuklisova-Murgasova M, Quaghebeur G, Rutherford MA, Hajnal JV, Schnabel JA. · 2012
Read Reconstruction of fetal brain MRI with intensity matching and complete outlier removal. by Kuklisova-Murgasova M, Quaghebeur G, Rutherford MA, Hajnal JV, Schnabel JA. (2012) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc4067058.txt.
Introduction
Maria Kuklisova-Murgasova, Gerardine Quaghebeur, Mary A. Rutherford, Joseph V. Hajnal, and Julia A. Schnabel presented a method for building a clear 3D image of a fetal brain from 2D MRI slices. Their paper appeared in Medical Image Analysis in 2012. A moving fetus blocks direct 3D scanning, so clinicians rely on stacks of slices.
Their reconstruction alternates between correcting slice motion and sharpening the image beyond normal slice resolution. Corrupted or misaligned slices are found and dropped automatically. A new step matches brightness between slices, which the authors show is essential for quality. They report excellent results on both clinical scans and optimized data. Better fetal brain images could support measures of growth and earlier spotting of abnormal development.
Catalog record
| CREATOR | Kuklisova-Murgasova M, Quaghebeur G, Rutherford MA, Hajnal JV, Schnabel JA. |
|---|---|
| YEAR | 2012 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Reconstruction of fetal brain MRI with intensity matching and complete outlier removal By Maria Kuklisova-Murgasova, Gerardine Quaghebeur, Mary A. Rutherford, Joseph V. Hajnal, Julia A. Schnabel Published in: Medical Image Analysis DOI: 10.1016/j.media.2012.07.004 ABSTRACT Graphical abstract Highlights ► A new method for reconstruction of 3D fetal brain MRI from 2D slices is proposed. ► The super-resolution reconstruction is interleaved with motion correction. ► Corrupted or misaligned slices are automatically excluded. ► Novel intensity matching is shown to be essential for quality of reconstruction. ► Excellent results for clinical and optimized data. INTRODUCTION Magnetic resonance imaging (MRI) of the fetal brain has received considerable attention in recent years due to its application in assessing fetal brain development. Clinically, it is used for qualitative assessment of fetal brain abnormalities. Its potential applications include detailed characterization of fetal brain development and identification of deviations in brain growth related to conditions such as intra-uterine growth restriction (IUGR) or preterm birth. Development of new quantitative biomarkers will facilitate better understanding of brain development and subsequent improved management of high-risk pregnancies (Rutherford et al., 2008; Limperopoulos and Clouchoux, 2009). Due to the long acquisition times of MRI, fetal motion presents a major challenge as 3D MR scanning sequences are not applicable in the presence of unpredictable and fast fetal motion. This has been addressed by the development of ultrafast sequences which are designed to acquire 2D slices and thus freeze the motion in time. The most common sequence used in clinical practice is single-shot fast spin echo (SSFSE) T2-weighted imaging with an acquisition time of 1 s per slice (Prayer et al., 2004). Stacks of thick slices are acquired in three orthogonal directions. The slice thickness is usually 3–4 mm to obtain good contrast-to-noise ratio. Such images provide excellent delineation of fetal brain anatomy and are suitable for qualitative assessment of brain abnormalities. However, compromised fetal brain development, e.g. during IUGR, often manifests as more subtle changes, which can only be assessed by detailed quantitative analysis. Studies on pre-term infants suggest that quantitative biomarkers, such as reduced cortical volume and sulcation index (Tolsa et al., 2004; Dubois et al., 2008) can be used to characterize IUGR. Quantitative analysis of brain development requires volumetric data and very recently, the development of methods for reconstruction of volumetric fetal brain MRI enabled the emergence of quantitative studies for fetal brain development (Damodaram et al., 2009a,b; Gholipour et al., 2011; Clouchoux et al., 2012; Rajagopalan et al., 2011a,b; Habas et al., 2012). In recent years, methods have been proposed to reconstruct a fetal brain volume from stacks of 2D slices (Studholme, 2011). Such reconstruction presents a number of challenges. Slices need to be automatically aligned to correct for the motion between acquisition of individual slices. Slices with motion artifacts need to be automatically excluded, and inconsistencies in intensity patterns resulting from the motion or acquisition settings, such as variable scaling of the slices and differential bias fields, need to be estimated and corrected for. Finally, the volume has to be reconstructed from irregularly sampled data. In the first approach proposed by Rousseau et al. (2006), the reconstruction using a weighted sum of Gaussian kernels which represent the point spread function (PSF) is interleaved with slice-to-volume registration and is applied to clinically acquired thick-slice data. If an interpolation approach is used to reconstruct a volume from thick slices, undesirable blurring is introduced into the reconstructed volume. Jiang et al. (2007) propose to acquire many thin slices (1 mm slice thickness) and use multilevel B-splines (Lee et al., 1997) to reconstruct the volume. One advantage of their approach is that the spline interpolation can be tuned to avoid blurring and the use of thin slices enhances this by achieving more isotropic resolution. As the region of interest is oversampled, averaging during reconstruction results in a significant reduction of noise, thus improving the signal-to-noise ratio. Additionally, thin slices can be aligned with significantly reduced target registration error (0.2 mm in simulated experiments). The method was further extended by a model of the motion of the fetus during acquisition to improve the robustness of the registration (Bertelsen et al., 2009). Kim et al. (2010) proposed an alternative motion correction approach, which replaces slice-to-volume registration with a multi-slice registration method constrained by a model of fetal motion. This removes the need for the computationally expensive reconstruction step during alignment of the slices. A modified Gaussian weighted reconstruction is proposed to reduce the blurring effect of thick slices. More recently, super-resolution methods (Milanfar, 2007; Greenspan, 2009) in combination with slice-to-volume registration have been proposed to reduce the blurring effects during reconstruction of thick-slice data (Gholipour and Warfield, 2009; Gholipour et al., 2010). During super-resolution reconstruction, the volume has to be regularized to prevent amplification of noise and registration error, as well as to fill the undersampled regions. To reduce the smoothing effects of regularization, adaptive regularisation techniques can be employed (Peyre, 2011). Rousseau et al. (2010, 2011) proposed to extend super-resolution reconstruction through total variation regularization or an edge-preserving regularization technique developed by Charbonnier et al. (1997). Kim (2011b) proposed an alternative non-iterative solution to selectively weight the Gaussian kernels to obtain a high-resolution volume. Early reconstruction approaches relied on manual exclusion of the motion corrupted slices. In more recent work, Kim et al. (2010) implemented a rejection of motion corrupted and misregistered slices based on an abnormally high mean squared intensity difference, but in several cases slices had to be manually marked for removal after processing. Gholipour et al. (2010) incorporated robust statistics based on the Huber function into super-resolution reconstruction to automatically reduce the weight of corrupted and misaligned voxels and slices. Recently, Kim et al. (2011a) showed that correction of bias field inconsistencies can improve the quality of the reconstructed volume, which is especially important for correct delineation of fetal lamination. In this paper we propose a comprehensive method for reconstruction of fetal brain MRI which builds and expands on previous super-resolution approaches (Gholipour and Warfield, 2009; Gholipour et al., 2010; Rousseau et al., 2010). For the first time, we unify all the necessary elements within a common framework to perform simultaneous super-resolution reconstruction, robust statistics and intensity matching to achieve high quality reconstruction. Our contribution compared to previous works is threefold: first, we propose to estimate the differential bias fields and slice-dependent scaling factors during super-resolution reconstruction; second, a novel robust statistics based on an EM framework is proposed to remove artifacts caused by motion-corrupted and misaligned data. Unlike the previously proposed work by Gholipour et al. (2010), the identified outliers are removed completely, thus further reducing the artifacts in the image. Third, we combine the robust super-resolution reconstruction with edge-preserving regularisation Charbonnier et al. (1997) which reduces blurring in the reconstructed images. The reconstruction is interleaved with slice-to-volume registration. We use normalized mutual information as a similarity measure due to its robustness to intensity artifacts and thus remove the need for a model of motion to achieve good quality alignment. Our method is evaluated using simulation of thick slice acquisition for volumetric neonatal data. We also propose a novel leave-one-out analysis to evaluate reconstruction results on real fetal data. Our results show that both intensity matching and the new robust statistics