Image fusion of multiple independent MRI brain slabs to cover a whole Dugong brain in a small-bore, high-field pre-clinical scanner
Simone Zanoni1, Kenneth WS Ashwell2, and Andre Bongers1

1Biological Resources Imaging Laboratory, The University of New South Wales, Sydney, Australia, 2Department of Anatomy, The University of New South Wales, Sydney, Australia


This study we report about MRI of a fixed Dugong brain in a small bore pre-clinical MRI scanner. To enable the scanning of the brain that exceeds the sensitive scanner dimensions we propose a multi-slab imaging approach that uses an optimized image post-processing pipeline to merge independent slabs into a continuous high-resolution high-contrast 3D volume. We demonstrate that using this imaging and post-processing approach it is feasible to investigate relatively large objects in a pre-clinical scanner and retain full 3D information with the full benefit from the superior high resolution imaging capabilities of a high field pre-clinical scanner.


The Dugong is an aquatic, herbivorous mammal and one of the four living Sirenia species that inhabits several territories throughout the Indo-West Pacific Ocean. The largest population of this endangered animal is found in the coastal waters of Western Australia. Currently, little is known about its brain structure1 and ethical constraints limit the availability of animals or specimens for MR studies. Here we report about technical aspects of the first MRI study on a fixed Dugong brain. High-resolution, high SNR MRI was performed using a small bore 9.4T pre-clinical scanner. Geometrical constraints in the scanner did not allow whole brain coverage and the MRI data were acquired in several independent slabs leading to distinct discontinuities between them due to coil profile and scan adjustments. We report about our approach to post-process the image data to merge the independent slabs into a cohesive and smooth high contrast 3D volume.


The brain was harvested from a Dugong that had died in captivity for unknown reasons, fixed in formaldehyde for approximatively 4 weeks and then immersed in 9g/l saline + 0.2%v/v GdDTPA for 48 hours prior to imaging2. Imaging was performed in a pre-clinical 9.4T Bruker BioSpec 94/20 MRI, with BGA-12S-HP gradients (Gmax=660mT/m, (dG/dt)max=4570Tm/s). To fit the relatively large brain into the scanner, a large ‘transmit-only’ volume RF-coil with a sensitive area of 86mm(ID)x90mm(length) was diverted from its intended use for transmission as well as signal reception. MRI comprised of a set of high resolution structural and DTI protocols to retrieve brain morphology and connectome. A 2D GRE method was used for high-resolution structural MRI with following major parameters: FoV=90x50mm, Matrix=512x256, Resolution=176x195μm, 40 slices/slab, Slice-thickness=500μm, TE=4.3ms, TR=690ms, FA=30°, 30 avg. To cover the whole brain length with the sensitive length, specimen was scanned in two independent (manually relocated) positions. Within each physical brain location 5 slabs were acquired with an overlap of each slab of 4 slices, to allow for individual manual slab adjustments and shimming. Acquisition adjustments caused substantial discontinuities between the independent image slabs.

To merge the discontinuous datasets and remove artefacts a post-processing pipeline was implemented and optimized (Fig. 2). First, noise is reduced using structure preserving filter provided in FSL (SUSAN)3,4. Subsequently, we used a histogram matching (HM)5 step to match the independent slabs to a reference histogram belonging to the slice with the highest CNR6. To cope with signal variations from inhomogeneities and coil signal drop off at the boundaries, a bias field correction module was implemented using the FSL FAST method7. This pipeline was run to merge all the slabs acquired without physical re-positioning of the sample for both positions independently. Both independently positioned merged slabs were then fused into one single dataset by co-registering their overlapping portions using a method by Collignon et al.8 with a contextual intensity normalisation.


Fig.3 shows the results of the image fusion of the discontinuous multi-slabs MR data using the proposed post-processing. Fig.4a) shows the in multiple independent slabs that were overlapped from the acquisition without further processing. It can be seen that the data have substantial discontinuities, due to the coil profile and inhomogeneities from independent shimming. HM levels out the most severe discontinuities (Fig.4c). Residual inhomogeneities within the slabs are then removed by the bias correction step (Fig4d).

The effect and inner workings of the most crucial step of the pipeline is shown in Fig.5, where the histograms from overlapping slices in different slabs are shown, before and after the procedure.


Coherent high resolution 3D datasets are most crucial for the neuroanatomical analysis of continuous structures and enable high-resolution structure segmentation and volumetry. This first study in a Dugong brain sample showed that it is possible to use a small bore pre-clinical scanner to produce continuous 3D volume data from relatively large brains that exceed the sensitive scanner volume by using a multi-slab acquisition approach and subsequent image fusion. The post-processing data fusion pipeline presented in this work was able to remove discontinuities and artefacts from independently acquired slabs and vastly reduced noise from the ‘raw’ data slabs while preserving tissue contrast. This substantially improves the ability of the observer to discriminate between and follow anatomical structures. And it is a crucial requirement for co-registration and overlay of functional and DTI information.


Our post-processing approach allows to merge MRI data with substantial discontinuities from acquisition in several independent slabs into a homogeneous 3D volume. This allows to scan brains that exceed the dimensions of a pre-clinical scanner and merge them into a coherent 3D dataset with high CNR.


The authors acknowledge the facilities, and the scientific and technical assistance of the National Imaging Facility at the UNSW Mark Wainwright Analytical Centre, Biological Imaging Resources Laboratory.


1. Pirlot P, Kamiya T. Qualitative and quantitative brain morphology in the Sirenian Dugong dugong Erxl. J Zool Syst Evol Res. 1985; 23(2):147-55.

2. Cherupalli S, Hardman CD, Bongers A, Ashwell KW. Magnetic Resonance Imaging of the Brain of a Monotreme, the Short-Beaked Echidna (Tachyglossus aculeatus). Brain Behav Evol. 2017; 89(4):233-48

3. Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. FSL. NeuroImage. 2012; 62:782-90.

4. Smith SM, Brady JM. SUSAN - a new approach to low level image processing. Int J Comput Vis. 1997; 23(1):45–78.

5. Rolland JP, Vo V, Bloss B, Abbey CK. Fast algorithms for histogram matching: Application to texture synthesis. J Electron Imaging. 2000; 9(1):39-46.

6. Brown RW, Haacke EM, Cheng YC, Thompson MR, Venkatesan R. Magnetic resonance imaging: physical principles and sequence design. John Wiley & Sons. 2014.

7. Zhang Y, Brady M, Smith S. Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximization algorithm. IEEE Trans Med Imag. 2001; 20(1):45-57.

8. Collignon A, Maes F, Delaere D, Vandermeulen D, Suetens P, Marchal G. Automated multi-modality image registration based on information theory. Inf Process Med Imaging. 1995; 3(6):263-74.


Figure 1. Superior view of the Dugong brain harvested from the animal and provided by the Taronga Zoo, Taronga Conservation Society Hospital, Sydney, Australia. Outer brain dimensions were measured (128mm length x 86mm width) and exceeded the maximum sensitive length of the pre-clinical coil.

Figure 2. Post-processing pipeline for dataset merging.

Figure 3. Coronal and axial sections of the structural Dugong brain before and after data merging with the proposed post-processing pipeline. a) Coronal slice of the ‘raw’ MRI data without any post-processing. Due to re-scanning and coil sensitivities the data slabs show significant bias and intensity variations causing discontinuities between the overlapping slabs; b) The same coronal slice after applying the complete pipeline. The procedure largely removes any bias. c) and d) Axial slice before and after the post-processing pipeline. The SNR and GM/WM CNR measurements below the slices show a significant reduction of image noise with well preserved structures.

Figure 4. Detailed view of MR data from one physical scan position after the individual pipeline processing steps. a) Brain slabs as acquired in one physical scan position; The discontinuities from RG adjustment and coil sensitivity variations are clearly visible. b) Dataset after noise reduction step using the SUSAN noise filtering algorithm; c) Dataset after cumulative distribution function (CDF) HM; This most crucial step removes most of the variability between slices. d) Dataset after bias Correction using FAST. This step removes residual inhomogeneities within individual slabs, reducing overall signal variations.

Figure 5. Effect of the CDF based histogram matching on an axial plane slice. Magnitude data of axial slices from reference and matching slices are shown in the upper row with corresponding histograms in the lower row. Left. Reference slice for matching. For matching, a maximum CNR slice was chosen to preserve maximum histogram spread. Middle. Matching slice from slab acquired in a different scan before histogram matching. Due to different individual adjustments histogram differs significantly from reference slice. Right. Matching slice after histogram matching step. Histogram matches the intensity profile of the reference slab closely.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)