Segmentation of the Gray and White Matter in the Human Spinal Cord
Neurological diseases like Multiple Sclerosis have an impact on the spinal cord. In this project, we study the effect of the disease on the different compartments of the spinal cord: gray matter and its surrounding white matter (GM/WM). To isolate these compartments, we segment magnetic resonance images of the spinal cord. Manual segmentation of such images is tedious and suffers from high intra- and interrater variability. We, therefore, aim at implementing deterministic and automatic segmentation algorithms to robustly and accurately mimic the manual segmentation process. We collaborate with the Division of Radiological Physics of the University Hospital Basel for developing image acquisition techniques that produce sharp images with high contrast between GM and WM, suitable for this segmentation task. Our developed segmentation algorithms we then use in collaboration with the Neurological Clinic and Polyclinic of the University Hospital Basel to calculate longitudinal atrophy rates of Multiple Sclerosis patients.
Axial MR image of the neck at C3 level acquired with the AMIRA sequence. On the right, we see a zoomed-in view on the spinal cord. The bright structures are the cervical arteries, the spinal fluid and the spinal cord gray matter.
Delineated in blue, we see a cross-sectional schematic of the spinal cord. It has a diameter of around 1 cm. On the MRI on the right, we can recognize the ellipsoidal shape of the spinal cord. The inner gray matter actually appears darker than the white matter, which is contrary to their names. This is, however, simply due to the contrast of this particular MR sequence.
Some interests of this work are individual atrophy rates on the separate gray matter and white matter compartments. Different MS patients may show different behavior, which might help to classify MS sub classes.
Different pros and cons for manual human segmentation and automatic computer-based segmentation. In this exemplary image, the computer algorithm failed to connect the left and the right side of the gray matter.
An energy formulation to segment gray matter on an image with gray matter contrast. This energy represents the so called continuous min-cut with an additive shape regularization term. We used this energy to derive an algorithm for gray matter segmentation for this project's first publication. Important information is, that the output of the algorithm is a segmentation which minimizes this energy.
This image shows exemplary images and segmentation attempts of the algorithm of our first publication. The middle column shows results without shape regularization and the right column the results of the proposed algorithm.
The MOLLI sequence: this image shows the timings of the inversion recovery MR sequence to acquire one part of the k space. The blue, green, and yellow curves represent approximate magnetization recovery curves of cerebrospinal fluid, gray matter, and white matter, respectively. After three 180° inversion pulses, in total 11 inversion images of the same slice are taken at different effective inversion times. These 11 images are shown below.
The MOLLI sequence: the 11 inversion images can be summed together (averaged). The resulting average image reveals remarkable gray matter-white matter tissue contrast.
The AMIRA sequence: this image shows the timings of the inversion recovery sequence and approximate magnetization recovery curves for the target tissues. At a resolution of 0.67 mm x 0.67 mm x 8 mm and a field of view of 128 mm x 128 mm, 17 inversion recovery preparations are needed to cover the full k space. This results in only 51 seconds for scanning one axial slice. Similar to the MOLLI sequence, the AMIRA sequence captures 8 inversion images of the same slice with different tissue contrast. These inversion images are shown at the bottom.
The AMIRA sequence: the 8 inversion images can be averaged differently. Averaging the first five images result in an image with remarkable gray matter-white matter contrast, averaging the last five images, improves contrast-to-noise ratios for spinal fluid.
Top row: different MR sequences used within the spinal cord gray matter segmentation challenge from 2015. Bottom row: different MR sequences developed in-house.
The AMIRA protocol: We use the AMIRA sequence to acquire 12 axial slices of the cervical spinal cord. The first slice is positioned such that the slice's lower edge coincides with the upper edge of the intervertebral disc between C2 and C3. Subsequent slices have an average slice distance of 4 mm and are aligned perpendicular to the spinal cord center line during acquisition. This results in a stack of slices in which we are able to segment the spinal cord gray matter and white matter. We extended the algorithm of this project's first publication to 3D and to other cervical levels and published it in a second publication at the American Journal of Neuroradiology (AJNR).
These boxplots show the optimal coefficients for combining the 8 inversion images of the AMIRA sequence to images with optimal contrast. At the bottom, we see images when using uniform averaging and the proposed non-uniform averaging. This analysis shows, that the uniform averaging techniques, we have already seen in the previous images, are already close-to-optimal.
Two exemplary images at C2 and C5 level acquired with the AMIRA sequence. In the top row, we see reproductions of histological drawing found in Grey's Anatomy.
The angulation experiment: this image shows different angulation of AMIRA slices. The effects of different angulation is shown in the next image.
Exemplary images acquired with different slice thicknesses and angulation. This experiment shows, that it is important to acquire the images perpendicular to the spinal cord center line - otherwise large partial volume effects for GM tissue are visible.
Exemplary AMIRA images of MS patients. The proposed manually designed segmentation pipeline fails at differentiating gray matter and hyper-intense lesion tissue, since the algorithm was not designed to do so.
Relative numbers of publications reported at sciencedirect.com in the field of medical image segmentation for the respective classes of algorithms. Deep learning establishes a new state of the art at around the year 2015, also in the field of medical image segmentation.
We used a recurrent neural network (RNN) with multi-dimensional gated recurrent units (MDGRU) to train GM-WM segmentation networks. We published the results at MICCAI CSI 2018. These three exemplary images demonstrate the accuracy of MDGRU, compared to manual human segmentations.
Further exemplary images of MS patients. In the future work, we try to differentiate GM, WM, healthy and lesion tissue. A first attempt is published at ISMRM 2020.
Manual Segmentation App: to acquire manual segmentations of the AMIRA images, we developed a browser-based application to display the various inversion images, as well as co-registered 3D T1 and T2 sequences, and to be able to simultaneously segment the images while switching between the different inversion images and their averages.
In another publication, we experimented with interpolating the slices acquired with the AMIRA protocol. Said protocol acquires images with a slice distance of around 4 mm, and, therefore, produces a stack of slices with anisotropic resolution. A finer through-slice resolution is beneficial for further processing. We developed a slice interpolation method that spline interpolates and optimizes the displacement fields between all subsequent slices in a global manner and simultaneously also spline interpolates the respective intensities along the registered spatial splines.
For more information about this project, please contact Prof Philippe Cattin.