Method, apparatus and system for identifying a specific part of a spine in an image

Abstract

A method, apparatus, and system for reliably identifying a specific part of a spine in an image of a human or animal body, includes the steps of determining one or more parts of the spine in the image, determining one or more discriminative parameters for each of the one or more parts of the spine in the image, the discriminative parameters relating to at least one anatomical property of each of the one or more parts of the spine, classifying the discriminative parameters of the one or more parts of the spine in the image, and identifying a specific part of the spine based on the classification of the discriminative parameters of the one or more parts of the spine in the image. An identified vertebra, in particular the T12 vertebra and/or its associated intervertebral discs, can be used advantageously as a starting point of powerful automatic spine labeling algorithms.

Claims

1. A method for identifying a specific part of a spine in an image of a human or animal body, the method comprising the steps of: determining one or more parts of the spine in the image; determining one or more discriminative parameters for each of the one or more parts of the spine in the image, the one or more discriminative parameters relating to at least one anatomical property of each of the one or more parts of the spine; classifying the one or more discriminative parameters of the one or more parts of the spine in the image; and identifying a specific part of the spine based on the classification of the one or more discriminative parameters of the one or more parts of the spine in the image; wherein the one or more parts of the spine in the image are determined on a basis of a contour image derived from the image by the following steps: detecting a spinal canal of the spine in the image; pruning away a frontal part of a rib cage from the image; calculating a frontal plane maximum intensity projection of the image; deriving a binary image from the frontal plane maximum intensity projection of the image by comparing pixel values of the maximum intensity projection with at least one bone threshold value; and deriving the contour image from the binary image.

2. The method according to claim 1, wherein the one or more discriminative parameters relate to angles between the spine and a rib contour at the one or more parts of the contour image.

3. The method according to claim 1, wherein the one or more parts of the spine in the image are determined based on the contour image by the following steps: smoothing the contour image; calculating local minima and local maxima in the contour image; determining vertebra centers on a basis of the local minima and/or the local maxima; establishing connections between the vertebra centers with corresponding local maxima; and determining the one or more discriminative parameters for each of the one or more parts of the spine in the image on a basis of the connections including lengths and/or angles of the connections.

4. An apparatus for identifying a specific part of a spine, in an image of a human or animal body, the apparatus comprising an image processor configured or programmed to: determine one or more parts of the spine in the image; determine one or more discriminative parameters for each of the one or more parts of the spine in the image, the one or more discriminative parameters relating to at least one anatomical property of each of the one or more parts of the spine; classify the one or more discriminative parameters of the one or more parts of the spine in the image; and identify a specific part of the spine based on the classification of the discriminative parameters of the one or more parts of the spine in the image; wherein the one or more parts of the spine in the image are determined on a basis of a contour image derived from the image by the image processor being further configured or programmed to perform the following steps: detect a spinal canal of the spine in the image; prune away a frontal part of a rib cage from the image; calculate a frontal plane maximum intensity projection of the image; derive a binary image from the maximum intensity projection of the image by comparing pixel values of the maximum intensity projection with at least one bone threshold value; and derive the contour image from the binary image.

5. A system for identifying a specific part of a spine in an image of a human or animal body, the system comprising: a computed tomography processor configured or programmed to acquire at least one image of at least a part of a human or animal body; and the apparatus according to claim 4.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 shows an example of an apparatus and a system according to a preferred embodiment of the invention.

(2) FIG. 2 shows examples of results of different processing steps of an input 3D CT scan.

(3) FIG. 3 shows an example of a smoothened rib contour (left) and a representation of a feature space (right), in which discriminative parameters, i.e. length and angle of connecting lines between centroids and rib/process tips of vertebrae, are plotted.

(4) FIG. 4 shows an example of a feature space in which discriminative parameters together with two linear SVMs learned from 29 CT images are depicted.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

(5) FIG. 1 shows an example of an apparatus 10 and a system according to a preferred embodiment of the invention. A medical image data set 11 comprising a plurality of images, in particular slice images, of a human or animal body is acquired by a medical imaging apparatus 12, in particular a computer tomography (CT) apparatus.

(6) The apparatus 10 comprises a control unit 13, e.g. a workstation or a personal computer (PC), to which the image data set 11 is fed. Preferably, the image data set 11 can be transferred from the medical imaging apparatus 12 to the control unit 13 via a data network 18 to which the control unit 13 is, at least temporarily, connected. For example, the data network 18 can be a local area network (LAN) or wireless LAN (WLAN) in a hospital environment or the internet.

(7) Preferably, the control unit 13 is configured to generate a slice view 15 and/or a volume reconstruction (not shown) of the image data set 11 on a display 14, e.g. a TFT screen of the workstation or PC, respectively.

(8) According to a preferred embodiment of the invention, the control unit 13 is designed to identify a specific part of a spine, in particular a specific vertebra, which can preferably serve as a starting point for an algorithm for labeling one or more further parts, in particular further vertebrae and/or intervertebral discs, of the spine in the image data set 11. In the example given in FIG. 1, a vertebra T12 corresponding to the 12th thoracic vertebra was identified and accordingly labeled in the axial slice view 15 of the image data set of a spine.

(9) In the following, the method and corresponding apparatus and system for identifying a specific vertebra, in particular the T12 vertebra, in the image data set 11 according to preferred embodiments of the invention will be elucidated in detail.

(10) The method relates to a, preferably machine learning-based, algorithm for reliable discrimination of the 12.sup.th thoracic vertebra T12 in lumbar-thoracic parts of CT scans by two discriminative features for separating the lumbar vertebrae from the thoracic ones. Preferably, the two discriminative features relate to the length and the angle of the attached transversal process or rib, both measured in a curved coronal projection of bones close to the spinal canal.

(11) Preferably, the image data set 11 is a three-dimensional (3D) data set and the discriminative parameters are determined on the basis of a two-dimensional (2D) data set which is derived from the original 3D data set 11. Based on spinal canal extraction, the frontal part of the rib cage is cropped out and the remaining bones attached to the spine are projected in a frontal curved maximum intensity manner.

(12) Post processing of this projection yields a single closed contour in 2D, further referred to as “rib contour” or “contour image”. Its interior is split into vertebral segments containing 2D projections of transverse processes or ribs. For each vertebral segment, a pair of discriminative features is computed and classified with a pre-computed support vector machine (SVM). The derivation of the 2D rib contour from the 3D volume data set 11 is detailed in the following.

(13) Detection of the spinal canal is a first logical step for any algorithm dealing with spine. Preferably, the spinal canal is detected or extracted by the method disclosed by F. Schulze, D. Major, and K. Bühler, Fast and memory efficient feature detection using multi-resolution probabilistic boosting trees, in Journal of WSCG, 19 (1):33-40, 2011, which is herewith incorporated into this patent application by reference.

(14) Once the spinal canal is extracted, the frontal part of the rib cage is pruned, i.e. image data relating to the frontal part of the rib cage are eliminated from the image data set 11.

(15) The remaining 3D data set is converted and/or displayed by a frontal, maximum intensity projection (MIP), wherein in the visualization plane the voxels, i.e. 3D image data, with maximum intensity that fall in the way of parallel rays traced from the viewpoint to the plane of projection are selected.

(16) Preferably, to account for the natural spine bending, eventual scoliosis, injury deformations or similar, the MIP is adopted in a curved planar reformation (CPR) manner according to the method disclosed by A. Kanitsar, D. Fleischmann, R. Wegenkittl, P. Felkel, and M. E. Gröller, CPR—Curved Planar Reformation, in IEEE Visualization 2002, pages 37-44, 2002, which is herewith incorporated into this patent application by reference.

(17) By the MIP, a 2D data set is derived from the original 3D data set. Therefore, from this point on, the rest of the algorithm according to a preferred embodiment of the invention refers to 2D data sets.

(18) Further, the MIP projection is thresholded to obtain bone pixels. The resulting binary mask is processed by elementary image processing operations resulting in several closed connected contours. The longest one, the rib contour, contains all the lumbar vertebrae and some lower thoracic vertebrae and is a subject for further analysis.

(19) The extraction of the rib contour from the 3D volume data set 11 is summarized in Algorithm 1 and illustrated by FIG. 2, which shows results of different processing steps, i.e. a maximum intensity projection of the input CT scan with a ground labeling overlay (left part), a curved planar projection of bones attached to the spine, with frontal ribs cropped out (middle part) and contours extracted from the mask, wherein the longest one is referred to as the rib contour 20 (right part).

(20) TABLE-US-00001 Algorithm 1 From 3D CT to 2D rib contour function GetRibContour(CT volume)      returns one contour  Detect spinal canal  Prune away the frontal part of rib cage  Curved maximum intensity projection         2D image  Bone threshold                   fig. 2 mid  Extract edges and connected contours        fig. 2 right  Return the longest contour (rib contour)         fig. 3 left end function

(21) In the following, the determination of vertebra segments and respective measures, in particular discriminative parameters, will be elucidated in detail.

(22) In order to achieve numerical stability of a subsequent analysis of the rib contour and to remove eventual outliers, the rib contour coordinates are smoothed by a Gaussian kernel. This also yields the rib contour's differentiable parameterization (x(t); y(t)), wherein x(t) and y(t) denote spatial coordinates of the contour with respect to continuous, real-valued parameter t.

(23) To extract vertebrae segments from the smoothed rib contour, a search for local minima and maxima of signed (planar) curvatures is performed

(24) $\begin{array}{cc}k=\frac{x^{\prime }{y}^{″}-y^{\prime }{x}^{″}}{{\left(x^{\mathrm{\prime 2}}+y^{\mathrm{\prime 2}}\right)}^{3/2}}& \left(1\right)\end{array}$

(25) In equation (1) k denotes a signed curvature of the rib contour.

(26) Assuming a counter-clockwise orientation of the curve, local maxima correspond to rib tips or process tips, respectively, and local minima correspond to approximate disk locations.

(27) This is summarized in Algorithm 2 and illustrated in FIG. 3 (left), which shows a smoothed rib contour 21 with local minima μ which approximate intervertebral disc positions and yield the segmentation (dotted lines) of the spine. For sake of clarity, only six minima μ of twenty-two minima in total are annotated with reference letter “μ”.

(28) Centroids of the segments annotated with reference letter “α” are connected to the two associated rib/process tips τ. For sake of clarity, only three centroids α of ten centroids in total (see symbols annotated with “L”, “X” and “T”) and six local maxima τ of twenty-four maxima in total are denoted with corresponding reference letters.

(29) The center α of each segment is connected with the two tips τ of the associated ribs/processes via a straight line δ. Two measures, which are also called “discriminative parameters” in the context of the invention, are calculated for the connecting lines δ. Preferably, the length and the angle enclosed by the left and right part the connecting lines δ are calculated in order to get a measure for distinguishing between Λ-like (+1), dash-like (0), i.e. straight, and eventual V-like (−1) connecting line δ.

(30) TABLE-US-00002 Algorithm 2 Vertebra segments and features function GetSegments(RibContour)         returns list of polylines   Smooth rib contour   Compute curvatures, minima and maxima         fig. 3 left, a, b   Connect corresponding minima         fig. 3 left, dotted lines   Vertebra centers from curvature minima         fig. 3 left, c   Connect centers with corresponding maxima      fig. 3 left, d   Return list of polylines end function

(31) In the following, the identification or discrimination of the T12 vertebra based on the two calculated measures will be elucidated in detail.

(32) It was found that lumbar polylines, i.e. connecting lines δ (see left part of FIG. 3), are short and their angles tend to cluster near zero, whereas the thoracic vertebrae are long with positive angles. The shape of the T12 ribs, however, can vary between that of lumbar processes and thoracic ribs. Thus, neither length nor angle is discriminative.

(33) In order to reliably discriminate the T12 vertebra from other vertebrae based on the length and angle of the connecting lines δ, two linear support vector machines (SVMs) are trained: the first SVM separating the lumbar segments {L1 . . . L5} from the thoracic ones {T7 . . . T12} and the second SVM separating the {T7 . . . T11} from {T12 . . . L5}.

(34) An SVM is a so-called classifier, which corresponds to a mathematical method for separating, i.e. classifying, an amount of objects into two or more classes. SVMs are supervised learning models with associated learning algorithms that analyze data and recognize patterns used for classification. A basic SVM takes a set of input data and predicts, for each given input, which of two possible classes forms the output, making it a non-probabilistic binary linear classifier. Given a set of training examples, each marked as belonging to one of two categories, an SVM training algorithm builds a model that assigns new examples into one category or the other. An SVM model is a representation of the examples as points in space, mapped so that the examples of the separate categories are divided by a clear gap that is as wide as possible. New examples are then mapped into that same space and predicted to belong to a category based on which side of the gap they fall on. Methods of training SVMs are disclosed by V. N. Vapnik, The nature of statistical learning theory, Springer-Verlag, New York, Inc., New York, N.Y., USA, 1995, which is herewith incorporated into this patent application by reference.

(35) FIG. 3 (right) shows the calculated length and angle of feature polylines, i.e. connecting lines δ, plotted in a two-dimensional feature space. The solid lines correspond to the two pre-learned support vector machines (SVM) that split the feature space into four sub-regions, i.e. lumbar region (“L”), T12 region (“X”), thoracic region (“T”) and blank. The dotted line connects the segment features in the anatomical order.

(36) FIG. 4 shows an example of a feature space in which discriminative parameters together with two linear SVMs learned from 29 CT images are depicted. As apparent from the example given in FIG. 4, there was no linear SVM (i.e. line) found that would clearly separate the T12 from the rest of the thoracic vertebrae, i.e., that would separate groups {T7 . . . T11} and {T12 . . . L5}, but a linear SVM was found to discriminate between the lumbar vertebrae segments {L1 . . . L5} from thoracic ones {T7 . . . T12}.

(37) To reliably find the T12 segment in an application, all available vertebral segments are first sorted in the bottom-to-top order by the y-coordinates of the centroids α and then classified as “L”, “X”, or “T”. In order to reliably identify the T12, there must be at least one segment classified as “L” and at least one segment classified either as “X” or “T”. The first segment following the chain of “L” segments is identified as T12.

(38) Examples of segment class strings in the bottom-to-top order with the final verdicts are given as follows:

(39) “LLLXTTT” T12 is the 4.sup.th segment

(40) “LLLLTTT” T12 is the 5.sup.th segment

(41) “LXXT” T12 is the 2.sup.nd segment

(42) “TTTT” no “L” segment available, i.e., T12 is uncertain

(43) “LLLL” only lumbar segments available, i.e., T12 is uncertain.