Patent Application
20080205724
The invention relates to a method for segmenting an anatomic structure in a multi-dimensional dataset comprising a plurality of temporally spaced cardiac images comprising data on a target matter and on an other matter.
The invention further relates to an apparatus for segmenting an anatomic structure in a multi-dimensional dataset comprising a plurality of temporally spaced cardiac images comprising data on a target matter and on an other matter.
The invention still further relates to a computer program for segmenting an anatomic structure in a multi-dimensional dataset comprising a plurality of temporally spaced cardiac images comprising data on a target matter and on an other matter.
An embodiment of the method as is set forth in the opening paragraph is known from U.S. Pat. No. 5,903,664. The known method is arranged to carry out an image segmentation step for purposes of identifying contiguous regions of the same target matter from suitable diagnostic images. In particular the known method is suited for segmenting the left ventricle from suitable diagnostic cardiac images. For this purpose in the known method a suitable region of interest in the cardiac images is determined under an operator supervision whereby an initial seed point within the envisaged region of interest is located. Also, an initial threshold for pixel or voxel classification is identified. Starting with a suitable initial image selected from the multi-dimensional dataset comprising temporally spaced cardiac images, points of the image within the region of interest are classified. Contiguous image elements having the same classification as the seed point and being connected to the seed point through image points all having the same classification are identified thus defining the thought segmented structure in the image.
It is a disadvantage of the known method that for enabling a segmentation of the thought anatomic structure, notably a ventricle in the heart an interaction with an operator is necessary whereby a threshold used for classification is defined. This results in a poor robustness of the known method with respect to both a user reproducibility and a segmentation accuracy. The former problem is explained by the fact that for the same multi-dimensional dataset different operators may select different thresholds. The latter problem is explained by the fact that an intensity of picture elements or volume elements representing fat in cardiac images is similar to those of blood leading in a poor differentiation between the ventricular tissue and fat tissue. This leads to inferior segmentation results.
It is an object of the invention to provide a method for segmenting an anatomic structure in a multi-dimensional dataset comprising a plurality of temporally spaced cardiac images comprising data on a target matter and on an other matter, whereby said method provides more accurate segmentation of the anatomic structure, notably a ventricle of the heart.
To this end the method according to the invention comprises the following steps:
performing a classification of cardiac images to distinguish between the target matter and the other matter yielding classified cardiac images comprising the target matter;
applying a thinning operator to the classified cardiac images yielding processed cardiac images comprising connected image components;
labeling different connected image components yielding respective labeled connected image components;
for each labeled connected image component compute a factor based on its volume variability with time;
segmenting the anatomic structure by selecting the connected image component with the factor meeting a pre-determined criterion.
The technical measure of the invention is based on the following insights:
i) the heart's ventricles exhibit coherence along all dimensions of a notably four-dimensional dataset. Specifically, within a cross-sectional slice space the core of the ventricle is substantially static for slices acquired for different longitudinal positions and for different temporal phases;
ii) the ventricles contract and expand significantly during the cardiac cycle, unlike the fat tissue.
Thus, based on these observation, the ventricles can be automatically distinguished among bright regions in the, for example, four-dimensional dataset, whereby regions are defined as connected clusters of bright pixels or voxels. For this purpose use can be made of per se known image processing techniques.
Therefore, in the first step of the method according to the invention performing a classification of cardiac images to distinguish between the target matter, notably blood and the other matter, notably non-blood is performed yielding classified cardiac images comprising substantially the target matter. This step can be enabled by using an automatic unsupervised binary voxel classification by computing the intensity histogram of the entire three-dimensional and temporal image. After this, a binary shareholding method is applied. An example of a suitable binary thresholding method is given in N. Otsu “A threshold selection method for gray-level histograms”, IEEE Transactions on System Man and Cybernetics, smc-9(1): 62-66. January, 1979. After the classified cardiac images are obtained, a suitable thinning operator is applied to the classified cardiac images yielding processed cardiac images comprising connected image components. The thinning operator is applied for the cross-sectional images, for example by utilizing “E”-morphological erosion steps with an 8-connected two-dimensional kernel, where E is preferably set to a value of 6.25 mm/voxel-X-size. Next, the step of labeling of connected image components is performed, whereby connectivity is determined using an 8-connected 4D kernel. Next, for each labeled connected image component a factor is computed, which is preferably based on a difference between a first volume of the connected image component and a second volume of the connected image component among all temporal phases of the cardiac images. Preferably, the first volume is set to a second largest volume and the second volume is set to a second smallest volume to ensure robust estimation of the volume variation in time. Finally, the anatomic structure is segmented by selecting the connected image component with the factor meeting a pre-determined criterion. Preferably, the pre-determined criterion is set as the largest value of said difference.
In an embodiment of the method according to the invention, the method further comprises a preparatory step of automatically computing a restrictive region of interest around the heart in the cardiac images of the multi-dimensional dataset. This technical measure ensures a substantial reduction of image information for segmentation purposes as parts of the image not belonging to the region of interest are neglected. Preferably, a method disclosed in C. A. Cocosco et al “Automatic cardiac region-of interest computation in cine 3D structural MRI”, Computer Assisted Radiology and Surgery (CARS), 2004 is used.
In a further embodiment of the method according to the invention, the method further comprises the steps of:
performing a region growing operation for the multi-dimensional dataset, whereby said region growing operation is being constrained by a parameter deduced from the classified cardiac images.
Preferably, this step is carried out using opening by reconstruction, for example implemented using morphological dilation with a 4-connected 2D-kernel in the cross-sectional slice plane as well as “D2”-dilation steps in the longitudinal direction using a 2-connected 1D-kernel, whereby the factor D2 is preferably set to 16 mm/voxel-Z-size.
In a still further embodiment of the method according to the invention, the method further comprises the following steps:
applying a thinning operator to the classified cardiac images yielding processed cardiac images comprising further connected image components;
labeling different further connected image components yielding respective labeled further connected image components;
for each labeled further connected image component compute a factor based on its volume variability in time;
segmenting the further anatomic structure by selecting the further connected image component with the value of said further factor meeting a further pre-determined criterion.
The resulting segmentation will advantageously comprise accurately segmented two anatomic structures, notably the left cardiac ventricle and the right cardiac ventricle.
In a still further embodiment of the method according to the invention the method still further comprises the step of segmenting a still further anatomic structure based on a comparison between the segmented anatomic structure and the segmented further anatomic structure.
This technical measure is based on the insight that the cardiac muscle surrounds the left ventricle and is partially bounded by the right ventricle. Thus, provided the left ventricle and the right ventricle are accurately segmented whereby fat tissue is robustly eliminated during the segmentation steps, the segmentation of the two ventricles provide a substantial segmentation of the cardiac muscle. The segmentation of the cardiac muscle is important for clinical studies aimed at wall thickness and motion analysis.
In a still further embodiment of the method according to the invention, the method further comprises the steps of:
computing a still further factor based on a relationship between the factor and the further factor;
comparing a value of the still further factor with a still further pre-determined criterion;
performing an automatic correction of a stack of cardiac images upon an event that the still further factor and the criterion inter-relate in a pre-determined way.
This technical measure is based on a further insight that due to imprecise scan planning or due to substantial axial heart motion, the basal short-axis transversal slice extends into the atria which may decrease the accuracy of ventricular segmentation. Further on, it is empirically determined that there is a reproducible indicator of such an event. Notably, when for the still further factor a ratio of the two largest values of the factor described above is selected, the criterion can be set to a simple numerical value. For example when the still further factor is given by F1/F2, whereby F1 is the largest value of the difference between a first volume of the connected image component and a second volume of the connected image component among all temporal phases of the cardiac images for the left ventricle and F2 is the same for the right ventricle, a correction of the stack of images is required when the ratio F1/F2 is greater than 4,0. The correction can be enabled by cropping the top Z slice in the four-dimensional image obtained after the thinning operator is applied to the classified image, then by repeating the labeling step, then growing the labeled components back into the top Z slice, preferably using an opening by reconstruction morphological operation. Concluding, the steps of region growing and segmenting are performed. This technical measure is particularly advantageous as it provides a fully automated means for image stack error detection and correction enabling a fully automated accurate and robust image segmentation method.
In a still further embodiment of the method according to the invention, the method further comprises the step of:
visualizing the at least any one of the segmented anatomic structure, the segmented further anatomic structure and the segmented still further anatomic on a display means.
It is considered to be advantageous to enable an investigation of the segmentation results by a user. An experienced user may detect minor segmentation failures, particularly when the image stack is erroneously prepared allowing an extension of the short-axis transversal slice into the atria. To correct for this, the user may manually mark a boundary between the left ventricle and the right ventricle, which can be enable by a convenient computer mouse action. In fact, in the usual situation where only the ejection fraction measurement is needed, it is sufficient to mark the boundary on two two-dimensional slices, one for end-diastole and one for end systole temporal phases. This feature will be explained in more detail with reference to
An apparatus according to the invention comprises:
an input for accessing the multi-dimensional dataset;
a computing means for:
i. performing a classification of cardiac images to distinguish between the target matter and the other matter yielding classified cardiac images comprising the target matter;
ii. applying a thinning operator to the classified cardiac images yielding processed cardiac images comprising connected image components;
iii. labeling different connected image components yielding respective labeled connected image components
iv. computing for each labeled connected image component a factor based on a its volume variability with time;
v. segmenting the anatomic structure by selecting the connected image component with the factor meeting a pre-determined criterion.
It is possible that the apparatus according to the invention is arranged as a working station, which may be arranged as a stand-alone device or may be connectable to a remote unit by means of suitable remote access facilities, like internet. Preferably, the apparatus according to the invention is further arranged with a suitable display unit for displaying the segmented anatomic structure. Advantageously, such configuration may be arranged as a viewing station, which is used for inspection of the segmentation results. Preferably, the apparatus according to the invention is further arranged with a suitable data acquisition unit for acquiring the multi-dimensional dataset. Preferred embodiments of the suitable data acquisition unit comprise a magnetic resonance imaging apparatus, a computer tomography unit, an X_ray device and an ultra-sonic probe. A preferable data acquisition mode for the magnetic resonance imaging unit is “balanced Fast Field Echo”, (bFFE). Further advantageous embodiment of the apparatus according to the invention will be discussed with reference to
A computer program according to the invention comprises instructions for causing the processor to carry out the steps of:
performing a classification of cardiac images to distinguish between the target matter and the other matter yielding classified cardiac images comprising the target matter;
applying a thinning operator to the classified cardiac images yielding processed cardiac images comprising connected image components;
labeling different connected image components yielding respective labeled connected image components;
for each labeled connected image component compute a factor based on its volume variability in time;
segmenting the anatomic structure by selecting the connected image component with the factor meeting a pre-determined criterion.
Preferably, the computer program according to the invention comprises a further instructions to cause the processor to carry out a further step of: automatically computing a restrictive region of interest around the heart in the cardiac images of the multi-dimensional dataset and/or a still further step of:
performing a region growing operation for a transversal slice plane, whereby said region growing operation is being constrained by a parameter deduced from the classified cardiac images.
Still preferably, the computer program according to the invention still further comprises instructions for causing the processor to carry out still further steps as are set forth with reference to claims 4, 5, 6, 7.
These and other aspects of the invention will be explained in further details with reference to figures.
Preferably, for reducing an amount of data to be processed at step 6 the image data is subjected to a restrictive region of interest determination, whereby substantially the cardiac tissue is left in the image, the background or other tissue information being suppressed or eliminated. Preferably, the method of automatic region of interest determination is carried out in accordance with C. A. Cocosco et al “Automatic cardiac region-of interest computation in cine 3D structural MRI”, Computer Assisted Radiology and Surgery (CARS), 2004.
At step 9 the classified cardiac images are selected in the transversal plane and subjected to a per se known image thinning operator, preferably by means of utilizing “E”-morphological erosion steps with an 8-connected two-dimensional kernel, where E is preferably set to a value of 6.25 mm/voxel-X-size. The resulting images comprise a plurality of connected image components which are further analyzed at step 14. It is noted that after the thinning step 9 a labeling step 11 is required, where different connected components in the multi-dimensional dataset are accordingly labeled. This step is preferably followed by a region growing step 13, which is constrained by binary threshold used at step 8b.
Next, for each connected image component a factor F is computed at step 14, which is based on a difference between a first volume of the connected image component and a second volume of the connected image component among all temporal phases of the cardiac images. Preferably, the first volume is set to a second largest volume and the second volume is set to a second smallest volume to ensure robust estimation of these volumes. Finally, the sought anatomic structure is segmented at step 16 by selecting the connected image component with the factor F meeting a pre-determined criterion. Preferably, the pre-determined criterion is set as the largest value of said difference. After this, the segmented anatomic structure, notably a ventricle, is stored in a suitable format at step 18.
The method 1 according to the invention may comprise additional advantageous steps to further increase the robustness of the segmentation result. Notably, for cases when the domain of cardiac image is inferiorly prepared, allowing the basal short-axis transversal slice to extend into the atria, the segmentation method according to the invention may experience some difficulties when separating left ventricle from the right ventricle. In order to eliminate this problem, in the method 1 according to the invention an automatic image domain correction step 17 is envisaged. This technical measure is based on an empirically determined fact that there is a reproducible indicator of such event. Notably, when for this indicator a ratio of the two largest respective values of the F-factor per ventricle is selected, the criterion can be set to a simple numerical value. For example when the ratio is given by F1/F2, whereby F1 is the largest value of the difference between a first volume of the connected image component and a second volume of the connected image component among all temporal phases of the cardiac images for the left ventricle and F2 is the same for the right ventricle, a correction of the stack of images is required when the ratio F1/F2 is greater than 4,0. The correction can be enabled by cropping the top Z slice in the four-dimensional image obtained after the thinning operator is applied to the classified image, then by repeating labeling step, then growing the labeled components back into the top Z slice, preferably using an opening by reconstruction morphological operation. Concluding, the steps of region growing and segmenting are performed. This technical measure is particularly advantageous as it provides a fully automated means for image stack error detection and correction enabling a fully automated accurate and robust image segmentation method.
In an alternative embodiment, after the segmentation step 16, the method proceeds to the step 19, whereby the segmentation results are displayed to the user on a suitable display means. Preferably the display mode comprises an overlay, notably in color, of the segmented anatomic structure on the cardiac images. In case the operator is satisfied with the results, the method stops at step 20. Alternatively, the operator indicates a boundary between the left ventricle and the right ventricle at step 22, after which this user-input is accepted at step 21 by a suitable per se known graphic user interface, after which the method returns to steps 14 and 16, which is carried out using a new geometric constrain, namely the boundary between the left and the right ventricle. It is noted that it is sufficient to mark said boundary only on two transversal sliced, one for an end-systole phase and one for the end-diastole phase. When the new segmentation is shown to the user at step 19 and the user is satisfied with the result, the method stops at step 20.
The core of the apparatus 30 is formed by a processor 34 which is arranged to operate the components of the apparatus 30, it being the input 32, the computing unit 35, the working memory 36, and the background storage unit 38. An example of a suitable processor 34 is a conventional microprocessor or signal processor, the background storage 38 (typically based on a hard disk) and working memory 36 (typically based on RAM). The background storage 38 can be used for storing suitable datasets (or parts of it) when not being processed, and for storing results of the image segmentation step, the step of determining respective volumes and F-factors, suitable criteria and thresholds as well as results of any other suitable intermediate or final computational steps. The working memory 36 typically holds the (parts of) dataset being processed and the results of the segmentation of the anatomic structure. The computing unit 35 preferably comprises a suitable number of executable subroutines 35a, 35b, 35c, 35d, 35e and 35f. The subroutine 35a is arranged to perform a classification of cardiac images to distinguish between the target matter, notably blood, and the other matter, notably fat tissue yielding classified cardiac images. The subroutine 35b is arranged to apply a thinning operator to the classified cardiac images yielding processed cardiac images comprising respective connected image components. The subroutine 35c is arranged to compute for each connected image component an F-factor based on a difference between a largest volume of the connected image component and the smallest volume of the connected image component. The subroutine 35d is arranged to perform suitable labeling of the connected image components. The subroutine 35e is arranged to segment the anatomic structure by selecting the connected image component with a maximum value of the F-factor.
Preferably, the computing unit 35 further comprises a subroutine 35f, arranged to compute a still further factor F′, based on a ratio between the respective F-factors for different anatomic structures, notably the left ventricle and the right ventricle. In case the F′ factor relates to a pre-determined criterion in a pre-determined way, this fact is signaled to the processor 34 as an event of the structure segmentation with reduced accuracy. In this case the processor 34 proceeds to a still further subroutine 35g, which is arranged to perform an automatic correction of the stack of cardiac images in accordance with the method of the invention discussed above.
The apparatus 30 according to the invention further comprises an overlay coder 37 arranged to produce a rendering of a suitable overlay of the original data with the results of the segmentation step. Preferably, the computed overlay is stored in a file 37a. Preferably, overlay coder 37, the computing unit 35 and the processor 34 are operable by a computer program 33, preferably stored in memory 38. An output 39 is used for outputting the results of the processing, like overlaid mage data representing the anatomy of the heart overlaid with the suitable rendering of the segmented structure. Further details are presented with reference to
Either of the data 42a, 42b, 42c or a suitable combination thereof is made available to a further input 45 of a suitable viewer 43. Preferably, the further input 45 comprises a suitable further processor arranged to operate a suitable interface using a program 46 adapted to control a user interface 48 so that an image of the anatomic data is suitably overlaid with the results of the segmentation step, notably with data 42a, 42b and/or 42c, thus yielding image portions 48a, 48b, 48c. Preferably, for user's convenience, the viewer 43 is provided with a high-resolution display means 47, the user interface being operable by means of a suitable interactive means 49, for example a mouse, a keyboard or any other suitable user's input device. Preferably, the user interface allows the user to interact with the image for purposes of marking a boundary between the left ventricle and the right ventricle, if necessary. Suitable graphic user input is translated into a geometric threshold by the computer program 46. This threshold is then provided to a computing means of the apparatus for a further iteration of the image segmentation step. This option allows for an accurate segmentation of the cardiac ventricles even in situations where the domain of input cardiac images is inferiorly prepared. Preferably, the apparatus 40 and the viewer 43 are arranged to form a viewing station 45a.
Preferably, for reducing an amount of data to be processed at step 56 the image data is subjected to a restrictive region of interest determination using a suitable computing algorithm, whereby substantially the cardiac tissue is left in the image, the background or other tissue information being suppressed or eliminated. Preferably, the method of automatic region of interest determination is carried out in accordance with C.A. Cocosco et al “Automatic cardiac region-of interest computation in cine 3D structural MRI”, Computer Assisted Radiology and Surgery (CARS), 2004.
At step 59 the classified cardiac images are selected in the transversal plane and subjected to a per se known image thinning operator, preferably by means of utilizing “E”-morphological erosion steps with an 8-connected two-dimensional kernel, where E is preferably set to a value of 6.25 mm/voxel-X-size. The resulting images comprise a plurality of connected image components which are further analyzed at step 64. It is noted that after the thinning step 59 a labeling step 61 is required, where different connected components in the multi-dimensional dataset are accordingly labeled using respecting computing routines. This step is preferably followed by a region growing algorithm at step 63, which is constrained by binary threshold used at step 58b.
Next, for each connected image component a factor F is computed at step 64, which is based on a difference between a first volume of the connected image component and a second volume of the connected image component among all temporal phases of the cardiac images. Preferably, the first volume is set to a second largest volume and the second volume is set to a second smallest volume to ensure robust estimation of these volumes. Finally, the thought anatomic structure is segmented at step 66 by selecting the connected image component with the factor F meeting a pre-determined criterion. Preferably, the pre-determined criterion is set as the largest value of said difference. After this, the segmented anatomic structure, notably a ventricle, is stored in a suitable format at step 68.
The computer program 50 according to the invention may comprise additional advantageous steps to further increase the robustness of the segmentation result. Notably, for cases when the domain of cardiac image is inferiorly prepared, allowing the basal short-axis transversal slice to extend into the atria, the segmentation method according to the invention may experience some difficulties when separating left ventricle from the right ventricle. In order to eliminate this problem, in the computer program 50 according to the invention an automatic image domain correction step 67 is envisaged. This technical measure is based on an empirically determined fact that there is a reproducible indicator of such event. Notably, when for this indicator a ratio of the two largest respective values of the F-factor per ventricle is selected, the criterion can be set to a simple numerical value. For example when the ratio is given by F1/F2, whereby F1 is the largest value of the difference between a first volume of the connected image component and a second volume of the connected image component among all temporal phases of the cardiac images for the left ventricle and F2 is the same for the right ventricle, a correction of the stack of images is required when the ratio F1/F2 is greater than 4,0. The correction can be enabled by cropping the top Z slice in the four-dimensional image obtained after the thinning operator is applied to the classified image, then by repeating labeling step, then growing the labeled components back into the top Z slice, preferably using an opening by reconstruction morphological operation. Concluding, the steps of region growing and segmenting are performed. This technical measure is particularly advantageous as it provides a fully automated means for image stack error detection and correction enabling a fully automated accurate and robust image segmentation method.
In an alternative embodiment, after the segmentation step 66, the method proceeds to the step 69, whereby the segmentation results are displayed to the user on a suitable display means using suitable graphic user interface routines. Preferably the display mode comprises an overlay, notably in color, of the segmented anatomic structure on the cardiac images. In case the operator is satisfied with the results, the computer program stops at step 70. Alternatively, the operator indicates a boundary between the left ventricle and the right ventricle at step 72, after which this user-input is accepted at step 71 by a suitable per se known graphic user interface subroutine, after which the computer program returns to the step of segmenting 74, which is carried out using a new geometric constrain, namely the boundary between the left and the right ventricle. It is noted that it is sufficient to mark said boundary only on two transversal sliced, one for an end-systole phase and one for the end-diastole phase. When the new segmentation is shown to the user at step 69 and the user is satisfied with the result, the computer program stops at step 70.
| Number | Date | Country | Kind |
|---|---|---|---|
| 05102864.5 | Apr 2005 | EP | regional |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/IB06/51112 | 4/11/2006 | WO | 00 | 10/11/2007 |
