JP2004097535A - Method for region segmentation of three-dimensional medical image data - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To perform time reduction of PCNN algorithm, and non-segmentation of an unnecessary region. <P>SOLUTION: In a method for setting a region to three-dimensional medical image data by applying a pulse combination neural network algorithm (PCNN), a seed point to a three-dimensional medical image is set based on input to an input device (S2), a threshold is obtained based on the seed point (S3 and S4), threshold processing is performed based on the value of the seed point and the value of a point adjacent to the seed point (S9) and a new seed point is decided based on the result of this threshold processing (S10). <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention generally relates to a method for segmenting (region dividing) medical three-dimensional image data (medical three-dimensional image data) obtained from a patient for diagnostic purposes.
[0002]
[Prior art]
Medical equipment with the processing speed and resolution capability that can obtain clinically useful medical three-dimensional image data is now becoming widely available in most large hospitals. For example, an ultra-high-speed MRI imaging technique such as spiral rotation CT and echo planar imaging, 3D ultrasound, or the like is used. At the same time, the computational power of the graphic computer has been continuously improved and the price has been reduced, so that more people can reconstruct, process and display such large medical three-dimensional image data.
[0003]
However, image processing by electronic means is the task of understanding and processing a particular image using millions of interconnected neurons (eg, the vast amount of data found in typical CT and MRI data). , The ability of the human brain to perform a task of dividing a region of interest (ROI) has not yet been reached.
[0004]
For this reason, several attempts have been made to mimic the function of the brain. One of the main examples is an artificial neural network (ANN). In ANN, various groups of units imitating neurons of a living body are connected to each other and weighted for these connections. These networks are usually composed of several stages (ie, inputs, outputs, and hidden), each of which requires training in its designed task. In recent years, a new type of ANN has become popular.
[0005]
This new type, called pulse-coupled neural network (PCNN), unlike conventional ANNs, consists of only one stage and does not require training. Its performance is determined by a set of parameters, not weighting. PCNN mainly extracts the basic conditions (edges, textures, segments) of the image and simplifies the original data and provides it to the next classification, quantification or display stage in the form of a binary or pulsed image.
[0006]
Other versions of PCNN have been proposed. This will be briefly described with reference to FIG. 1 based on the Abrahamson model. That is, each voxel (or pixel) in the original data is associated with one neuron in the PCNN neuron plane. This neuron is connected to a range of neurons on the same plane and to adjacent neurons on an adjacent plane corresponding to adjacent images in the volume.
[0007]
The central voxel is called the supply input (F) and the sum of adjacent neurons is called the connected input (L). The value of F is set to voxel intensity without any change, while the value of L is weighted. This weight is gradually attenuated. The intermediate value (internal activity) (U) is obtained by adding the L value to the F value, and then the value of U is compared with the threshold value T. If the value of U is greater than T, the neuron is considered "fire" and its ID is set to "1". Otherwise, it is set to zero.
[0008]
Each of the peripherally fired neurons (voxels) is set as a central voxel, and the weight value of the sum of the adjacent neurons (voxels) is added to the voxel value, and further compared with a threshold value T. As described above, the firing voxels are successively chained to expand the cluster. This process is repeated until no peripheral neurons fire at the current threshold.
[0009]
At time t = 0, the threshold is set to a very high value so that firing of the neuron never occurs. At time t = 1 (first time step), the threshold T is attenuated according to a user-specified decay constant, and its value is compared to the U values of all neurons. If there are no more neurons that can fire during this time step, time is advanced. T is attenuated again and the process is repeated.
[0010]
At each time step, a group of firing neurons forms a cluster. The value of the cluster is set to the number of the time step at which the corresponding neuron fired. The PCNN algorithm ends when the time step advances a specified number of times or when all neurons (voxels or pixels) in the image data have fired.
[0011]
A process flow of region division (segmentation) based on PCNN will be described in order with reference to FIG. Medical three-dimensional image data (volume data) is loaded, and PCNN parameters are set. The key parameters are the connection radius (r) (ie, the number of connection inputs connected to each neuron), the connection strength (β) (ie, the weight added to the sum of all L values), and the decay of L and T. A constant and the number of time steps. At time step # 1, each neuron is compared to T and the above process is repeated until no more neurons fire. The time advances to the next time step. T is attenuated and all neurons are compared to T. Neurons that fired during the previous time step are ignored. Repeated the specified number of times. A result consisting of all clusters is output.
[0012]
This conventional PCNN algorithm has the following two basic problems.
First, it takes time. At each time step, threshold processing is performed on all voxels in the volume. For a volume size obtained at the typical 512x512x500 of current imaging devices, it may take several minutes for the PCNN algorithm to finish. This is not acceptable if PCNN is to be applied in a clinical setting.
[0013]
Next, an unnecessary part is extracted. Typically, the region of interest consists of a small subset of voxels grouped into a particular cluster. However, since the PCNN extracts an area of no interest, the user needs to select a target cluster.
[0014]
[Problems to be solved by the invention]
An object of the present invention is to reduce the time of the PCNN algorithm and to not divide unnecessary areas.
[0015]
[Means for Solving the Problems]
According to a first aspect of the present invention, in a method for setting an area in medical three-dimensional image data by applying a pulse-coupled neural network algorithm (PCNN), a seed point is added to the medical three-dimensional image based on an input to an input device. Is set, a threshold value is determined based on the seed point, threshold processing is performed based on the value of the seed point and the value of a point adjacent to the seed point, and based on the result of the threshold processing. To determine a new seed point.
According to a second aspect of the present invention, there is provided a method for setting a region in medical three-dimensional image data by applying a pulse-coupled neural network algorithm (PCNN), wherein the two-dimensional image data including a region of interest is extracted from the medical three-dimensional image data. A first step of creating and displaying a two-dimensional image based on the two-dimensional image data; a second step of setting a seed voxel in the two-dimensional image based on an input to an input device; A third step of determining a threshold value based on the threshold value; and a fourth step of comparing the value determined from the values of the seed voxel and its surrounding voxels with the threshold value to determine firing of the seed voxel; A fifth step of executing the third and fourth steps using the voxel fired by the determination as a new seed voxel.
[0016]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, a method for dividing a medical three-dimensional image data region using a pulse-coupled neural network (PCNN) algorithm according to the present invention will be described with reference to the drawings. The algorithm of the pulse-coupled neural network (PCNN) is typically realized by software. However, the present invention is not limited to practical use by software, and may be realized by hardware designed according to the algorithm of the pulse-coupled neural network (PCNN).
[0017]
The medical three-dimensional image data (also referred to as volume data) is an ultra-high-speed MRI apparatus (magnetic resonance imaging apparatus) such as spiral rotation CT (helical X-ray computed tomography apparatus), echo planar imaging, etc. It is generated by an ultrasonic diagnostic apparatus and a gamma camera (SPECT, PET).
[0018]
First, as an outline of the present embodiment, a pulse-coupled neural network (PCNN) determines whether or not to include a voxel in the same cluster by associating not only the luminance of the voxel but also the luminance of the surrounding voxels with the luminance. It is based on doing. That is, the decision whether to include or exclude a voxel from the cluster is made by comprehensively characterizing the voxel and its surrounding voxels in a manner similar to how brain cells process information.
[0019]
Conventionally, all voxels constituting a volume are examined several times under various conditions. Therefore, it takes much time to implement the PCNN algorithm by software. Also, many unimportant clusters are split together with important clusters of interest. As a result, important clusters of high interest may be buried in many unimportant clusters.
[0020]
This embodiment is devised in order to shorten the processing time of the PCNN algorithm, and to divide (extract) only important clusters of interest by eliminating the division of many unimportant clusters. I have.
[0021]
The PCNN algorithm according to the present embodiment has a function of, when a user designates a desired position in a region of interest, setting one voxel corresponding to the position as a seed voxel (a search start point).
[0022]
Then, the threshold value is initialized so that the seed voxel fires. Specifically, the threshold value is attenuated at a predetermined time constant. In this case, firing means that the value of the seed voxel exceeds a threshold.
[0023]
It is checked whether the voxel adjacent to the seed voxel fires at the threshold value. The value of an adjacent voxel is added to a value obtained by multiplying the sum of neighboring voxels or a value derived from the sum by a weight (connection weight). The added value is compared with a threshold value, and when the added value exceeds the threshold value, it is determined that the adjacent voxel has “fired”. The firing voxels are classified into clusters of seed voxels.
[0024]
This thresholding process is repeated individually for all voxels adjacent to the seed voxel. The same threshold processing is repeated with the fired adjacent voxel as a new seed voxel. In this way, "firing" is checked in a chain around the seed voxel.
[0025]
Finally, when all of the adjacent voxels no longer fire, the PCNN algorithm starting from the seed voxel ends. A similar process may be repeated with other voxels set as seeds to obtain other clusters.
[0026]
In this embodiment, by restricting the scope of the PCNN algorithm to a small sub-volume, the PCNN process does not repeat several times with the time step attenuating the entire volume, but instead of repeating it several times, as in the prior art. Executed only once. For large volumes (512x512x500), this leads to a significant reduction in processing time. Another important advantage of seed PCNN segmentation is that only important seed clusters can be split without splitting irrelevant clusters.
[0027]
In addition, in order to promote elimination of irrelevant clusters, in order to eliminate irrelevant clusters, in particular, clusters composed of voxels having a luminance value higher than the region of interest, the present embodiment sets an upper limit value of luminance. I do. Voxels having a luminance value higher than this upper limit value are excluded from the target of the segmentation in advance.
[0028]
Hereinafter, a specific description will be given. FIG. 2 shows a procedure of a method for shortening the time required to extract a region of interest from medical three-dimensional image data, that is, the PCNN processing time, and improving the performance of PCNN segmentation by not extracting irrelevant clusters. I have.
[0029]
First, volume data is loaded (S1). In step S1, a plurality of parameters used in the PCNN algorithm are set. The setting of this parameter is typically performed by a user via an input device. However, the parameters may be set automatically to optimal values.
[0030]
Next, as shown in FIG. 5, an arbitrary section including a region of interest is set by the user with respect to the volume, a two-dimensional image of the section is created and displayed from the volume data, and the interest of the two-dimensional image is displayed. A seed is set in the area (S2). That is, by moving the cursor to a desired position in the region of interest in the two-dimensional image and clicking the mouse button, the voxel in the region of interest corresponding to that position is set as an initial seed voxel. The coordinates of the voxel (seed voxel) at the cursor position and the corresponding luminance value (SV = seed value) are stored.
[0031]
After the initial seed voxel selection of the PCNN, the threshold T is attenuated (S3). At the threshold value, it is checked whether or not the brightness value (SV = seed value) of the seed voxel can be fired as it is (S4). The threshold T is attenuated once or several times according to a predetermined time constant until the seed voxel can fire. An ID (time step number, here, 1) is assigned to the seed voxel and stored (S5). The stored voxels are firing voxels. Firing means that the intermediate value U exceeds a threshold.
[0032]
After S6 (described later), the position of the firing voxel is read (S8), and it is checked whether or not each voxel adjacent to the firing voxel fires (S9). That is, as shown in FIG. 3, adjacent voxels (three voxels having luminance values of 6, 5, and 7 in FIG. 4) adjacent to the seed voxel (in FIG. 4, voxels having a luminance value of 9) are connected with each other as a center. The sum of a plurality of voxels distributed around within a distance of radius r, or a connected input (L) derived from the sum is weighted by a concatenation weight β, and the value of the seed voxel, that is, the supply input (F) is β Are added to obtain a work intermediate value (internal activity) (U), and then the value of U is compared with an initially set threshold value T. If the value of U is larger than T, the adjacent voxel is regarded as “fired”, given “1”, and is stacked (saved) (S10). Otherwise, it is set to "0". In the example of FIG. 4, a voxel having a luminance value of 7 is ignited.
[0033]
Further, the voxel having the luminance value of 7 that has fired is set as a new seed voxel, and voxels around the voxel having the luminance value of 7 that have already been fired are omitted. Are examined individually. The sum of a plurality of voxels distributed around within a distance of a connection radius r around each of them, or a connection input (L) derived from the sum is weighted with a connection weight β, and the value of the seed voxel, that is, the supply input The work input value (internal activity) (U) is obtained by adding the connection input L weighted by β to (F), and then the value of U is compared with the threshold value T that is initially set. If the value of U is larger than T, the neuron is regarded as "ignited", given "1", and is stacked (saved) (S10). Otherwise, it is set to “0”.
[0034]
Such a chained firing check is repeated until no adjacent voxel is fired (S6, S8, S9, S10). If no neighboring voxels fire, the result is output in S7, and the PCNN algorithm ends. Conventionally, the PCNN algorithm is repeated many times while the threshold value T is attenuated until all the voxels are fired. However, in the present embodiment, the threshold value T is fixed to the initially set value, the firing of the adjacent voxel is determined based on the threshold value, and the determined voxel is replaced with the seed voxel. The PCNN algorithm terminates when no adjacent voxel fires any more and the stack becomes empty, that is, when no firing voxel ID is stored. .
[0035]
FIG. 6 is a schematic diagram showing, for comparison, results obtained by a conventional PCNN-based segmentation algorithm using the volume shown in FIG. 3 as an input. A dotted arrow overlappingly displayed on the right image indicates that the entire volume is being processed. Along with the ROI cluster (abdominal aorta), different clusters corresponding to each group of related voxels in adjacent organs (liver, kidney, etc.) are displayed. In some cases, a cluster other than the ROI cluster may block the display of the ROI so that the spine overlaps most of the aorta when viewed from the back. Therefore, additional processing steps are required to separate the ROI from other clusters.
[0036]
On the other hand, FIG. 7 shows a result obtained by the seed-based PCNN segmentation algorithm according to the present embodiment. Dotted arrows overlaid on the ROI indicate that only the ROI is being processed. Thus, the only resulting cluster is that ROI, and no additional steps are required to separate that ROI from other clusters.
[0037]
FIG. 8 shows that in the present embodiment, when temporally continuous image data is sequentially applied to the PCNN, the PCNN application range is limited to a range (sub-volume) including the ROI extracted in the previous image. Is shown. By limiting the PCNN application range to the sub-volume including the ROI extracted in the previous image, the processing can be further shortened.
[0038]
As described above, in the present embodiment, the segmentation process ends when all the voxels adjacent to the seed voxel no longer fire. The result of the segmentation consists of a single ROI cluster that is output to a file for display and / or analysis. Therefore, irrelevant clusters, such as those obtained with the conventional PCNN segmentation method, are not included, which facilitates analysis and interpretation of the results. The above described process can be repeated to split multiple ROIs. The final output in this case consists of the aggregate of all ROI clusters in a single volume. Accelerating PCNN-based segmentation allows for fine tuning of PCNN parameters for the purpose of improving the segmentation function. This can be very useful in dealing with low contrast data sets such as those obtained with ultrasound imaging devices. Naturally, many modifications of the present embodiment are possible from the viewpoint of the above contents. That is, when applied to segmentation of a medical three-dimensional image data set, this is a method for improving the speed performance and ROI extraction capability of a pulse-coupled neural network, and comprises the following.
[0039]
FIG. 9 is a flowchart of the seed-based PCNN parameter optimization processing described above. For this optimization, for example, a two-dimensional image is cut out from a volume (image selection), and the preprocessing and the trial area of the PCNN algorithm are limited from the two-dimensional image (area selection, area analysis), and the preprocessing and the PCNN algorithm are performed. This is performed by observing the respective results while approximating the parameters to the optimum values while repeating the preprocessing and the PCNN algorithm with the parameter change.
[0040]
I didn't know how to analyze the area and how to optimize the parameters from the results. Also, I didn't know exactly what the pre-processing was like (parameters couldn't go so far).
[0041]
Please add or modify the following sentence.
[0042]
As shown in FIG. 11, as a setting method of the target area, an auto and a manual are prepared. In the case of the auto, when a seed is designated by a user, in addition to a method of searching for a periphery based on a luminance value, A radial search is performed centering on the seed, an edge (outermost) point where the luminance or the luminance gradient satisfies a predetermined condition is extracted, and the points at the ends are connected to draw a contour of the analysis area. The analysis area is set through the repetition of connection involving addition, deletion, and movement of points.
[0043]
As pre-processing, for example, as shown in FIG. 10, a seed voxel value is corrected by a factor (correction value) so that more than half of the surrounding voxels are fired, and this is performed individually for all voxels in the target area. Based on the luminance distribution of voxels in the region, the PCNN algorithm sets an upper limit of the target luminance value, extracts only voxels below the upper limit from the entire volume, and performs PCNN only on the extracted voxels. It is processed by an algorithm.
[0044]
The PCNN parameters include a connection radius r and a connection weight β. After setting, preprocessing and PCNN segmentation are performed with the parameters, and the results are displayed. The user checks and, if necessary, resets the parameters.
[0045]
The setting of the PCNN parameters is provided by a graphical user interface (GUI), for example, as shown in FIG. The GUI is configured in an easy-to-understand manner such as the size, sharpness, and persistence of the area as a PCNN result, and each is provided with a slider button to assist in setting PCNN parameters that are difficult for the user to understand. For example, the size and the sharpness are determined by the correlation between the connection radius r and the connection weight β. For each of the size and the sharpness, a combination of the connection radius r and the connection weight β is associated with the set value (the size and the sharpness) in advance.
[0046]
Since this parameter optimization process is performed for the entire volume, it takes a long time. For this reason, as shown in FIG. 13, it is preferable to optimize the parameters limited to a part (sub-volume) of the volume.
[0047]
(Modification)
The present invention is not limited to the above-described embodiment, and can be implemented in various forms without departing from the spirit of the invention at the stage of implementation. Furthermore, the above-described embodiment includes various stages, and various inventions can be extracted by appropriately combining a plurality of disclosed components. For example, some components may be deleted from all the components shown in the embodiment.
[0048]
【The invention's effect】
ADVANTAGE OF THE INVENTION According to this invention, the setting of a favorable region of interest by PCNN can be performed at high speed.
[Brief description of the drawings]
FIG. 1 is a flowchart showing a procedure of a conventional PCNN.
FIG. 2 is a flowchart illustrating a procedure of a PCNN according to an embodiment of the present invention.
FIG. 3 is an explanatory diagram of threshold processing in a firing determination step of peripheral neurons (peripheral voxels) in FIG. 2;
FIG. 4 is an explanatory diagram of a threshold decay step in FIG. 2;
FIG. 5 is a supplementary explanatory diagram of the seed input step of FIG. 2;
FIG. 6 is a supplementary diagram of the conventional PCNN algorithm.
FIG. 7 is a view showing an example of a result output of FIG. 2;
FIG. 8 is a diagram showing a chain of PCNN application ranges when temporally continuous image data is sequentially applied to a PCNN in the embodiment.
FIG. 9 is a flowchart illustrating a procedure for adjusting a PCNN parameter in the embodiment.
FIG. 10 is an explanatory diagram of the preprocessing in FIG. 9;
FIG. 11 is a flowchart showing a procedure of region selection in FIG. 9;
FIG. 12 is a view showing an example of a GUI for setting PCNN parameters in FIG. 9;
FIG. 13 is a flowchart showing another procedure for adjusting PCNN parameters in the embodiment.
[Explanation of symbols]
S1 ... input step of medical three-dimensional image data (volume data) and PCNN parameters;
S2: seed input step,
S3: threshold attenuation step,
S4: seed firing determination step,
S5: Firing neuron stack step,
S6: Stack empty judgment step,
S7: Result output step,
S8: Neuron ID reading step,
S9: peripheral neuron firing determination step,
S10: Firing neuron stack step.

Claims (10)

In a method of setting a region in medical three-dimensional image data by applying a pulse-coupled neural network algorithm (PCNN),
Setting a seed point in the medical three-dimensional image based on an input to the input device;
Determining a threshold based on the seed point;
A threshold value process is performed based on the value of the seed point and a value of a point adjacent to the previous seed point, and a new seed point is determined based on a result of the threshold process. Data area division method. In a method of setting a region in medical three-dimensional image data by applying a pulse-coupled neural network algorithm (PCNN),
A first step of creating two-dimensional image data including a region of interest from the medical three-dimensional image data and displaying a two-dimensional image based on the two-dimensional image data;
A second step of setting a seed voxel in the two-dimensional image based on an input to an input device;
A third step of determining a threshold based on the seed voxels;
A fourth step of comparing the value obtained from the values of the seed voxel and its surrounding voxels with the threshold to determine firing of the seed voxel;
A fifth step of executing the third and fourth steps using the voxel fired by the determination as a new seed voxel, and a fifth method of segmenting medical three-dimensional image data. 3. The method according to claim 2, wherein, in the fourth step, a distance between the peripheral voxel and the seed voxel is set via the input device. 4. The method of dividing medical three-dimensional image data according to claim 3, wherein a slider button is provided by a graphical user interface in the form of the size of the divided region for inputting the distance. 3. The method according to claim 2, wherein, in the fourth step, the value of the seed voxel is weighted with a sum of values of the surrounding voxels or a value derived from the sum, and the sum is compared with the threshold value. A method for segmenting medical three-dimensional image data as described. 6. The method according to claim 5, wherein the weight is set via the input device. 7. The method according to claim 6, wherein a slider button is provided by a graphical user interface in the form of the size of the divided area for inputting the weight. 3. The medical three-dimensional image data area according to claim 2, wherein an area including a point specified via the input device or an area defined by an outline is analyzed, and parameters of the PCNN are set according to a result of the analysis. Split method. The method according to claim 1, further comprising: repeating a trial of the PCNN while changing a parameter of the PCNN for an area including a point specified by the input device or an area defined by a contour, so as to approximate the parameter of the PCNN to an optimal value. 2. The method for segmenting medical three-dimensional image data according to item 2. 9. The method according to claim 8, wherein the analysis area is set based on an intensity gradient of a predetermined number of lines radially set around the point.

Images