JP2008503294A - Identification method, computer program, and computer program device - Google Patents

Abstract

In a method for detecting a junction between a lesion and a wall in a CT scan image, a boundary (B) of the wall is determined with respect to an internal space (L), and singular points (c ₁ , c ₂ ) along the boundary are determined. ) And selecting one singularity on either side of the lesion as the junction between the wall and the lesion is provided. The singular point may be a point of maximum curvature at the boundary and a transition point between a straight line and a curved portion. The singularity accepts the first and second seed points (p ₁ , p ₂ ) on both sides of the lesion and moves the seed point relative to the boundary if they are not already on the boundary, It may be selected by searching for the nearest singularity. The seed point may be determined by moving the determined junction point (j ₁ , j ₂ ) from the adjacent slice image to the current slice image.

Description

  In particular, the present invention is a computer-aided tomography (CT) scan image, but is not limited thereto, such as a wall and lesion, such as a polyp in contact with a pulmonary nodule or colon in contact with a pleura in an image scanned in advance from a human body or animal It is related with the method of detecting a junction part. The present invention includes software and apparatus for performing the method.

Detecting lesions and suspicious sites at an early stage of cancer is considered the most effective way to improve survival. Detection of pulmonary nodules and detection of polyps is part of the challenge to be achieved in the medical image processing field, although it is quite difficult.
Computer-aided techniques have been proposed to identify regions of interest containing nodules in CT scan images, which separate nodules from surrounding objects such as blood vessels or lung walls, and calculate physical properties of nodules. And / or provide automatic diagnosis of nodules. In a fully automated method, all steps are performed without radiologist intervention, but one or more steps may require input from the radiologist, which is described as a semi-automated method. May be.
US Patent Application Publication No. 2003/099384

  Detecting the size or extent of the lesion is important for accurate diagnosis, but detecting the extent of a pulmonary nodule that touches the pleura or separating the nodule from the pleura is similar in CT scan images It is difficult to have strength. Similarly, it is difficult to detect the boundary between the polyp and the colon wall.

US 2003/099384 and US 2003/099389 disclose methods and forms for detecting pleural nodules by morphologically closing the contours of small nodules. A deformable surface model is disclosed for a nodule that is larger than the structural element used to close the contour.
International Publication No. 03/010102 pamphlet and Med. Phys. Volume 29, Issue 11 (published November 2002), pages 2552-2558, Gurcan M et al., “Detection of lung nodules in computed tomography images of the chest: computer "Preliminary evaluation of Assisted Diagnostic System (Lung Nodule Detection on Thoracic Computed Tomography Images: Preliminary Evaluation of a Computer-aided Diagnostics System)" Disclosed is a method for detecting pleural nodules using a local indentation search in the vicinity of the pleura, in which those along the boundary are disclosed. The ratio of the distance between the is searched to be larger than the linear distance between only two points amount that exceeds a predetermined threshold.

According to the present invention, in a method for detecting a junction between a lesion and a wall in a scan image, a boundary with respect to an internal space is determined, a singular point along the boundary is identified, and the wall and the lesion are A method is provided for selecting one singular point on either side of the lesion as the junction between. The singular point may be a point of maximum curvature at the boundary and / or a transition point between a straight line and a curved portion. The singularity accepts first and second seed points on either side of the lesion and moves the seed point relative to the boundary if they are not already on the boundary, and the singularity closest to the seed point It may be selected by searching for points. The seed point may be determined by moving the selected junction point from the adjacent slice image to the current slice image.
The advantage of this method is that the contact point between the lesion and the pleura can be accurately determined in three dimensions, and thus the range of the lesion can be determined more precisely.
The location of the junction point may be refined in various ways, and in some cases may not exactly match the determined singularity.
The present invention is preferably implemented using a computer and extends to software that performs the method.

CT Images Each embodiment is performed on a series of CT scan slice images obtained from a CT scan using a human or animal specimen. Each slice image is a two-dimensional digital grayscale image of X-ray absorption in the scan area. The characteristics of the slice image depend on the CT scanner used. For example, a high resolution multi-slice CT scanner generates an image with a resolution of 0.5 to 0.6 mm / pixel in the x and y directions (ie, on each slice image plane). Each pixel may have a 32-bit grayscale resolution. The intensity value of each pixel is usually expressed in Hounds field units (HU). A series of slice images may be spaced apart in the z direction (i.e., scan separation axis), for example, at intervals of 0.75 to 2.5 mm. Thus, the scanned image is a three-dimensional (3D) grayscale image and has a total size that depends on the area and number of scanned slice images.
The present invention is not limited to a particular scanning scheme. It can be applied to electron beam CT (EBCT), multi-detector or helical scan, or any scheme that generates 2D or 3D images representing X-ray absorption.
As shown in FIG. 1, the scan image is generated by the computer 4, and the computer receives the scan data from the scanner 2 and constructs the scan image. The scanned image is stored as one or a plurality of electronic files, and is stored on a storage medium 6 such as a fixed type or a removable disk. The scanned image may be processed by the computer 4 to identify the extent of the pulmonary nodule, or the scanned image may be transferred to another computer 8 that executes software for processing the image as described below. The image processing software may be stored on a carrier such as a removable disk or may be downloaded over a network.

Nodules Attached to the Lung Wall In certain embodiments, the boundary between the nodule and the lung wall is designed to be detected. The surrounding pulmonary nodules are often somewhat joined to the surface of the pleura (pressed against the thoracic boundary around the lungs). Nodules share their surface with the pleura to a considerable extent. For this reason, showing the boundary between the pleura and the nodule is a difficult problem. This complexity is reduced by using two seed points as inputs to identify the perimeter of the chest wall at opposite ends of the nodule. The seed point may be selected by a user such as a radiologist or may be determined automatically. The theory and examples of the boundary drawing method are shown below.

In a first embodiment, the method comprises the following steps outlined with reference to FIG.
1) Input seed points (step 12).
2) Perform rough separation to separate the lungs from surrounding tissues and determine the boundaries (step 14).
3) A singular point on the contour is determined (step 16).
4) The following processing is executed for each slice image (starting from the start slice image).
a. If each of the current points is not within the starting slice image and is within the tissue, the current point is moved to the boundary (step 18). If this process fails, the process ends.
b. If each point is in the lung space, move it towards the other point until it hits the boundary (step 20). Finish if it fails.
c. The nearest singularity along the contour is searched for the two current points (step 22).
d. The point is refined (step 24).
e. A nodule contour is obtained in the current slice image (step 26).
f. Check for nodule overlap in the previous slice image. Finish if there is no overlap.
g. Two points are mapped to the next slice image (step 30).
5) Repeat the process for consecutive slice images in both directions from the start slice image.

Input seed point (step 12)
The user examines one or more slice images of the CT scan image and visually identifies potential pleural nodules. Depending on the user input device, the user selects as two seed points both sides of a possible pleural nodule in one of the slice images. The seed points need only be on either side of the pleural nodule and need not be exactly on the boundary.
Alternatively, the user may use an input device to draw a box or other graphic around a possible pleural nodule. The crossing point between the figure and the pleura may be selected as two seed points.
As another option, the scanned image may be pre-processed to identify possible pleural nodules, and the seed point may be entered at this pre-processing stage.

Division (step 14)
A partial image consisting of multiple slice images (for example, 30) above and below the slice image formation image including the seed point is selected, and the image is separated into lung space and pleura (including any nodules attached to the pleura) In this way, the rough division is performed. Since the intensity of the lung space is much lower than that of the pleura and nodules, the segmentation may be performed by applying a predetermined threshold to the image. Objects above that threshold but not contiguous to the pleura are not considered by this algorithm.
In addition, in order to carry out with higher reproducibility, an optimum center slice image may be searched and the optimum center slice image may be set as the center of a partial image. For example, a window of a 60 pixel × 60 pixel × 10 slice image is determined, and its center is set as a midpoint between two seed points. A point having the maximum intensity in the window is searched, and a corresponding slice image including the point having the maximum intensity is determined as a central slice image of the partial image.
The division is further improved by performing an adaptive division after the initial coarse division. In adaptive segmentation, the lung space obtained by the coarse segmentation is expanded using distance transformation to obtain a mask including the lung space and a peripheral edge (edge) of the lung space. The magnification factor may be 50%. Such adaptive segmentation is performed with a threshold obtained from only the intensity values in the mask.
The result of the segmentation is a binary map of slice images, including pixels that have one of the two values and indicate the lung space, and pixels that have the other binary value and indicate the surrounding tissue. Pixels on the boundary between the lung space and the surrounding tissue are labeled as boundary pixels.
As an example, FIG. 3a shows a scanned image of a lung slice and FIG. 3b shows the result of the segmentation. FIG. 3b is a binary image in which the lung space is white (binary number 1) and the surrounding tissue is black (binary number 0). FIG. 3c is white and shows the boundary pixels obtained by segmentation.

Determination of singularity (step 16)
As a preliminary step to detect the junction between the nodule and the pleura, the singularity of the boundary is determined. Each section of the boundary is identified as a recess, protrusion, or straight line. A singular point is either the maximum concave or convex curvature point, or a transition point between concave, convex, or straight states. In other words, it is between the straight part and the convex part or the concave part, or directly between the convex part and the concave part. The singularity is important because it can indicate the boundary between the pleura and the nodule.
A particular method for determining singularities is described. A boundary contour map is input. This may be, for example, a binary array in which the position in the array is indicated by the x and y coordinates of each pixel. The boundary pixel is then flagged with a binary “1”. In other words, it is a boundary bitmap.
Next, the boundary pixel is selected as the starting point. This can be any boundary pixel, for example, the first boundary pixel is searched with a raster scan of the bitmap (left to right, top to bottom).

The algorithm then records an angle between successive adjacent boundary pixels and moves along the boundary. Since the pixels of each slice image are arranged in 2D Cartesian coordinates, only 8 angles are possible. For each pixel, adjacent pixels are represented by a grid as shown in Table 1.
The angle between each pixel and adjacent pixels can be coded as shown in Table 2.
The codes correspond to the angles shown in Table 3.
A bitmap is a one-dimensional code consisting of a code indicating the angle between successive adjacent boundary pixels in a loop around the boundary, for example (1, 1, 1, 2, 1, 3, 3...) Converted to chain code with vector. These are absolute angles with respect to the downward direction.
The chain code is then converted to the boundary curvature angle at each boundary pixel. This is done by taking the chain code around each boundary pixel and weighting so that the chain code for the boundary pixels near the current boundary pixel is more heavily weighted. Note that the curvature angle is relative to the boundary rather than the absolute angle so that the previous chain code of the current pixel on the loop is subtracted from that next of the current pixel on the loop. It has become.

As an example, the curvature angle for boundary pixel i is calculated as shown in the equation. Where m _j is a weighting factor according to a Gaussian distribution, cc _{i + j} is a chain code for pixel (i + j), is negative for negative j, and n is, for example, n = 5, The numbers taking into account each direction around the boundary are shown.
The result is a one-dimensional vector that indicates the curvature angle of the boundary at each boundary pixel with respect to neighboring pixels in both directions around the boundary. The effect of the weighting filter is to smooth the angle for the exact variance value of the chain code.
Next, the maximum and minimum curvature angles are identified. For example, neighborhood comparison methods may be used, where each angle is compared to a predetermined number (eg, three) of angles in the vector.
In order to identify a transition point between a straight line and a concave or convex part, an angular range corresponding to a “straight line” is defined. For example, all angles of 180 ° ± f may be classified as straight lines.
Figures 4a to 4c show the singularities for f = 5, 15, and 30 °, respectively. The transition between the concave maximum and the straight line and the concave portion is indicated by a square, and the transition between the convex maximum and the straight line and the convex portion is indicated by a circle. FIG. 4a shows that if f is set too low, a large number of singularities are produced that show only a gradual transition at the boundary curvature. The value of f may be preset by software and / or adjusted by the user.

Searching for nodule orientation Before detecting the nodule, it is first determined in which direction the nodule extends in the lung for the _two seed points p ₁ and p ₂ . In one embodiment, as illustrated in FIG. 5, a vector v is set between _two points p ₁ and p ₂ and an orthogonal vector o extends from the midpoint of the vector v to both sides. The direction of o toward the lung space L indicates the direction of the lung, that is, the direction in which the pleural nodule N extends from the pleura P into the lung.
After the initial slice image, the nodule direction is determined relative to the seed points p ₁ and p ₂ from the detected nodule direction in the previous slice image.

Move to space (step 18)
For each slice image, seed points p ₁ and p ₂ first determine if they are on boundary B, otherwise they are moved to the appropriate point on the boundary. In the initial slice image, the user need not place the seed points p ₁ and p ₂ exactly on the boundary. After moving the seed points onto the boundary and changing their z-coordinate to move to the next slice image, the seed points may no longer be on the boundary of the new slice image.
If either seed point p ₁ or p ₂ is in pleura P, that point is moved onto the boundary. The direction and distance of movement may be limited so that it does not move away from one singular point, so as to indicate the interface between the nodule and the pleura. Specifically, the angle of the interface is not changed greatly from one slice image to the next. Figure 3 illustrates one way to move a seed point from tissue to the boundary.
First, it is determined whether an arbitrary lung space L exists in a window (for example, 20 mm × 20 mm) centered on the moving seed point. If not present, an error condition may be indicated. If it exists, the following steps are executed as shown in FIG.
Draw two lines l ₁ and l ₂ from the points p ₁ and p _{2 in} the tissue in the direction o of the nodule N by 45 and 135 degrees from the vector v by 10 units each.
Generate a minimum-sized rectangle R from the moving seed point to the ends of the _two lines l ₁ , l ₂ (if _{one of} the points p ₁ , p ₂ is not in the tissue, such a pair of lines Note that you only need to have
In the rectangle R, a point closest to the seed point on the boundary B is searched.
・ If the moving distance is less than 10 mm, it moves to a point on the boundary.
This method places an appropriate limit on the angle at which the vector v can rotate and the distance that the seed point can move.

Integration of two points (step 20)
If each of the two seed points p ₁ and p ₂ is in the lung space L, that point is moved towards the other point until it hits a boundary B as shown in FIG. When the moved points are within a predetermined distance (for example, 2 mm) from each other, this indicates that there is only a space between the points, and the processing ends.

Search for nearest (nearest) singularity (step 22)
Both seed points p ₁ and p ₂ should now be on boundary B. They are then moved relative to the nearest singularity along boundary B according to certain conditions intended to avoid selecting inappropriate singularities, as described in more detail below.
For each seed point, the closest singular point in both directions along the boundary B is determined. If a singular point in one direction along the boundary is significantly closer than a singular point in the other direction, the seed point is moved to that close singular point. For example, if the distance (size) between the seed point and the closest singular point in one direction is greater than a certain threshold, the seed point is moved to the nearest singular point in the other direction. The threshold may be a function of the distance between the closest of the two closest singularities in each direction and the seed point, such as twice the distance.
If the closest singularities in each direction are approximately equidistant, one singularity is selected based on the amount of lung space on the straight line between the seed point and each singularity. For example, as shown in FIG. 9, if the difference in distance between the _two closest singular points c ₁ and c ₂ along the boundary B and the seed point p ₁ is less than the threshold, the two points from the seed point p ₁ A straight line l ₃ to each of the nearest singular points c ₁ and c ₂ is generated, and the ratio of the line length passing through the lung space L is determined. If the ratio exceeds a predetermined threshold (for example, 20%), the singular point is regarded as an inappropriate selection, and if the ratio to the nearest singular point c _{2 in} the other direction is less than the threshold, the seed point is the point. Moved to. If both singularities have a percentage greater than the threshold, the seed point is not moved to either singularity.

Movement when points are too close In some cases, one of the seed points may be relatively far from the best singularity, so both seed points are the same when searching for the nearest singularity It may be moved to a singular point or to a singular point that is quite close. This can occur when the seed point is mapped from the previous slice image, especially when fine segmentation movement is applied as described below. If the distance between two seed points along the boundary is less than a threshold distance such as 5 mm, the seed point closer to the singular point is moved to that singular point, while the initial distance between the seed points is The other seed point is moved along the boundary to the nearest singular point away from the other seed point, provided that the distance between the two seed points does not increase beyond a predetermined threshold. In other words, the distance between the seed points after moving away from one side should not exceed twice the distance before the movement.

When there is no singular point in the vicinity When there is no singular point in the vicinity of one of the seed points, the seed point may be moved to the closest singular point with respect to the other seed point. This can be prevented by avoiding that one seed point is moved along the boundary beyond a certain percentage, such as 50% of the distance between the seed points before movement, in which case it is separated from any singular point. Seed points are not moved.

Precision division (step 24)
After searching for the nearest singular point using the coarsely divided boundary, the movement of the seed point is refined by re-evaluating the boundary B in the vicinity of the seed point using fine division according to the local contrast. Good. For example, the threshold value for determining whether each point is in the foreground or the background may be determined based on the intensity of each point in a mask composed of a predetermined range of local points surrounding the point. The threshold may be selected such that the average of the intensity centers above and below the threshold is equal to the threshold or deviates from the threshold by a predetermined amount.
Next, singular points are determined on the finely divided boundaries. The seed point may be moved to the nearest singular point on the finely divided boundary according to certain conditions, as illustrated below.
• The movement of the seed point to the finely divided boundary is always away from the nodule direction.
・ The distance traveled is less than a predetermined threshold. The threshold is set according to how far the seed point has moved from the initial point in the slice to the final point using coarse partitioning.
The angle at which the line between the seed points rotates when moving to a singular point on a finely divided boundary is less than a predetermined angle such as 40 degrees.
The coarsely divided boundary may not be near the finely divided boundary of a nodule with a strong partial volume effect. In this case, where the closest singularity may leave the nodule, only movement away from the nodule is allowed, so the correct singularity is not searched. This means that if the seed point is inside the lung compared to a finely divided boundary, the seed points will move towards each other until they reach the boundary and search for the nearest singular point on the finely divided boundary. Can be solved.
The effect of precision segmentation is to eliminate the influence of substances surrounding the nodule, such as infected tissue that obscure the junction of the nodule to the lung membrane. Coarse subdivision usually includes such materials as part of the boundary, while fine subdivision reclassifies such materials as background.

When there are a plurality of boundaries, the lung L is divided into two or more parts by coarse division, and the seed points are located in different parts as illustrated in FIG. 10, for example. This can happen when the blood vessel extends across the lung in the current slice image. In addition, including a blood vessel as the nodule N portion may be an obstacle to determining the nodule N range.
In order to overcome this problem, if the seed point is moved to boundary B, it is determined whether to form a part of the boundary that is not in communication. If so determined, as shown in FIG. 11, the change in the distance between the two boundaries between the two seed points in the nodal direction o is calculated. FIG. 12 shows a graph as an example of the change in distance. A point g showing the largest change in the gradient of the graph is obtained, and that point is incorporated into the boundary B. In this way, the nodule N is separated from the blood vessel.

Contour correction For example, as shown in FIG. 13a, if the edge of the boundary between the detected junctions has gone too far, the singularity may not be in the correct position for the contact of the nodule to the lung wall. This occurs, for example, due to gradual angular changes at the junction.
To avoid this problem, a contour correction function may be used, and the junction points are readjusted so that correct regions can be extracted, as shown in FIGS. 13b to 13d. First, a line is drawn between the joining points. If the line intersects the boundary, move one junction along the boundary toward the other until it no longer intersects the boundary on that side. Then, the other junction is moved toward the first junction until it does not cross the boundary.
In some cases, especially when the nodule is small, the line between the junctions may not intersect the boundary even if the edges are overrun, as shown in FIG. 14a. In this case, as shown in FIG. 14c, by changing the line in small increments until the line crosses the boundary, the joint point is simultaneously moved along the boundary as shown in FIG. 14b. Then, as shown in FIG. 14d, the above-described contour correction is advanced until the line does not cross the boundary.

Acquisition of Nodal Boundary At this stage, two junction points have been determined, indicating the point along boundary B where nodule N is joined to pleura P. These junction points can be used to estimate the boundary of the nodule N in the slice image. As a first approximation, a straight line connecting the junction points is used to indicate the junction between the nodule N and the pleura P. Alternatively, a curve having that estimated from the curvature of the pleura surrounding the nodule N may be used.
Next, the range of the nodule N is determined. In one embodiment, it is assumed that nodule N is not in contact with or near any major object other than pleura P. As shown in FIG. 15, the division result of step 14 is used to identify all foreground regions beyond the line connecting the junction points when forming the nodule N portion.
In other embodiments, a fuzzy contrast region expansion method is used to identify the nodule N range. First, the range of the nodule N is estimated by a division process based on the local threshold. A local mask is used to obtain a threshold intensity value for each point. Thus, the threshold is sensitive to local contrast and low contrast nodules can be distinguished from the background. The nodule region N obtained by fine division may include a “hole”, ie, a background surrounded by foreground pixels. Holes are created based on the sensitivity of fine division to small local differences in contrast. A suitable hole filling algorithm is used to fill such holes, in other words, convert them to foreground pixels.
Next, the pixel with the highest intensity in the nodule region N is taken as a seed point, and binary region expansion is performed from that seed point to identify a continuous foreground region F. This region is expanded using a distance transformation method with an expansion factor to obtain a mask M that includes both foreground and background. The expansion factor may be 50%. However, as shown in FIG. 16, the mask M does not include pixels on the other side of the line connecting the junction points, but includes a projection of the line beyond the junction points. Next, the foreground pixels divided by the fine division process that are not continuous foreground regions F are removed from the mask M. The background pixels of mask M are labeled as background region B.
The mean and standard deviation of the intensity and gradient distribution of the foreground and background regions F, B are determined. An average μ _B of background intensity, an average μ _{F of} foreground intensity, a standard deviation σ _{B of} background intensity, and a standard deviation σ _{F of} foreground intensity are calculated. The parameter ε is estimated by counting the number of foreground intensity standard deviations σ _F where the seed point is away from the average of the background intensity μ _B , which is obtained as a measure of contrast and the fuzzy map described below. Used subsequently to build.
The fuzzy object extraction method is used to establish a 2D map of the fuzzy continuity of each pixel in the mask M relative to the seed point. A fuzzy affinity function between adjacent pixels is defined and fuzzy between each pixel and seed point by searching for affinity along the path between the pixel and seed point. Continuity is obtained.
Fuzzy continuity between two points (not necessarily adjacent) is obtained by considering the path with the strongest affinity between the two points. The path with the strongest affinity is selected as the best path, and the intensity of each path is equal to that of the weakest affinity of neighboring points along the path. The intensity between two adjacent points is the affinity between those points.
The affinity between two points is a measure of the probability that they belong to the same object. This probability is a function of proximity (ie, Euclidean distance) and the closeness of image features between those pixels.

The general model for fuzzy affinity is
μ _k = (c, d) = h (μ _a (c, d), f (c), f (d), c, d)
Given in. Here, h is a scalar value in the range [0, 1], c and d are the positions in the image of two pixels, and f (i) is the intensity of spell i. μ _a is an adjacency function based on the distance between two pixels and is given by an equation.

To simplify the shift independence, it is given by an equation when c ≠ d, and further μ _k (c, c) = 1.
It is.

Here, the subscript “i” indicates the calculation regarding the intensity, and “gk” indicates the calculation regarding the gradient value in each of the appropriate directions (which may be x, y, and z). ω _i and ω _g are weighted values of the free parameters whose sum is 1. Values of 0.9 for ω _i and 0.1 for ω _g have been chosen to make the intensity closeness effective. The fuzzy affinity used for the current function is shown in the equation.

Here, m _i , S _i , _mg and S _g are Gaussian distribution parameters for intensity and gradient. These may be pre-defined or estimated from a small area around the seed point as described below.
The average and standard deviation σ of intensity and gradient are calculated over all points in the mask M. Parameters related to the gradient are calculated separately in three directions (x, y and z). The corresponding σ is calculated based on the difference between the maximum and minimum slopes.
The statistical calculation is described in detail below. The parameter m _i is the starting intensity, and m _gx and m _gy are the average of the gradients in the x and y directions, respectively. The parameters S _gx and S _gy are standard deviations of gradients in the respective directions.
The standard deviation (S _d ) appearing in the affinity equation affects the formation of the fuzzy map and therefore the nodule boundary determination. If it is too large, the affinity curve will be relatively flat. As a result, the background area has a higher affinity and the expanding area is over-divided. Conversely, if it is too small, the shape of the affinity curve will be narrow, the foreground will have a low affinity with the seed point, and the result will only be under-divided. Ideally, the curve should have a minimal affinity for the background seed point but a finite degree of dispersion. Thus, the affinity Gaussian curve is limited or expanded by changing the standard deviation to achieve ideal dispersion.
A fuzzy map is then created by searching for the fuzzy continuity value of each pixel relative to the seed point. A fuzzy map may be viewed as an enhanced image that shows how strongly that pixel is associated with a seed point.
Next, starting from the seed point, the fuzzy map is subjected to region expansion based on contrast. As each pixel is added to the region during region expansion, the peripheral contrast is calculated between the interior and exterior of the region. The peripheral contrast of the region may be defined as the difference between the average gray level of the inner boundary and the average gray level of the current boundary.
In each iteration of contrast-based region expansion, one pixel is selected from the current boundary and added to the current region. The pixels preferentially selected at the current boundary are determined based on the intensity and the distance to the center of the current region. As each pixel is added to the region during region expansion, the peripheral contrast is calculated between the interior and exterior of the region. The peripheral contrast of the region may be defined as the difference between the average gray level of the inner boundary and the average gray level of the current boundary. The peripheral contrast at each stage is recorded and the area expansion continues until the mask M is filled. The highest peripheral contrast value obtained during region expansion is chosen as indicating the optimal region with a boundary corresponding to that of nodule N.

Map to next slice image (step 28)
After the range of nodule N in the current slice image is determined, it is determined whether the nodule N in the current slice image overlaps the range of nodule N determined in the previous adjacent slice image. This may be done in various ways, for example, determining a circular “nucleus” of a predetermined radius surrounding the center of the nodule N in each slice image, and whether or not the adjacent slice images overlap. May be done by determining whether there are any pixels having the same x and y coordinates between the two nuclei. If there is no overlap, that is, no nodule is detected in the current slice image, this may indicate that the boundary of nodule N in the current direction from the starting slice image has been reached. If the algorithm proceeds in only one direction from the starting slice image, it proceeds in the other direction. If nodule boundary detection is performed in both directions, the algorithm ends and outputs the range of detected nodules in each slice image in which nodule N is detected.
If there is an overlap with the previous slice image, the algorithm proceeds to the next slice image and, as shown in FIG. 17, the junction points j ₁ and j ₂ of the current slice image are set to the two seed points p of the next slice image. _1, is used as the p _2. For example, the seed point of the new slice image has the same x and y coordinates as the junction point of the previous slice image.

Results FIG. 18 shows the range of nodules N detected in the scanned image of the sample.

Application to Large Intestine The above-described embodiment can also be applied to detection of polyps in a CT scan image of the large intestine. Polyps are always attached to the colon wall, which has a similar intensity in CT scan images. The shape of the polyp is slightly different from that of the pulmonary nodule, but the joint between the colon wall and the polyp has similar geometric characteristics, and the embodiment should be used to detect the joint Can do.
The present invention is applicable to the detection of other abnormal tumors attached to the walls of the internal space of the human body and / or animals.

Other Embodiments The embodiments described above together with the examples do not limit the scope of the present invention. Various modifications may be envisaged within the scope of the claims. As is clear from the above discussion, the method can be performed on a 2D image consisting of a single CT slice image or a 3D image consisting of successive CT slice images.

FIG. 2 is a schematic diagram showing a CT computer and a remote computer that processes image data from the scanner. It is a flowchart of the algorithm in embodiment of this invention. (A) is a diagram showing a single slice image of a CT scan image of a lung, (b) is a diagram showing a single slice image of a CT scan image of a lung region obtained by segmentation, (c) is a lung region It is a figure which shows the single slice image of CT scan image of no boundary. FIG. 5 shows a singular point detected on the lung boundary by one of three different thresholds. It is a figure which shows the seed point of the both sides of a nodule. It is a figure explaining the method of moving a seed point from a pleura to a boundary. It is a figure explaining the method of moving a seed point from a lung space to a boundary. It is a figure explaining the method of moving a seed point to the nearest singular point. It is a figure explaining the method of determining the best singular point when it exists in a substantially equidistance. It is a figure which shows the example by which the lung was divided | segmented into the some boundary. FIG. 4 shows where the junction between boundaries is determined. Fig. 4 is a graph showing how a bond is determined and showing the width against the distance on the midline between two boundaries. It is a figure which shows the method of correct | amending the edge which the nodule went too much. It is a figure which shows the method of correct | amending the edge which passed too much when the detected edge has exceeded the nodule. It is a figure which shows the range of the detected nodule. It is a figure explaining another method of detecting the range of a nodule. It is a figure explaining mapping a junction point to the seed point of the next slice image. It is a figure which shows the range of the nodule detected in a series of continuous slice images.

Claims (29)

In a method for identifying a junction between a lesion in a CT (Computer Aided Tomography) scan image and a wall adjacent to the internal space,
a. Determine the boundary between the wall and the interior space;
b. Determining one or more singularities along the boundary;
c. Selecting first and second junction points on the boundary on both sides of the lesion based on the location of at least one singular point on each side of the lesion;
As a result, the junction between the lesion and the wall is identified by the junction.   The method of claim 1, wherein the singularity includes at least one curvature maximum along the boundary.   3. The method according to claim 1 or 2, wherein the singular point includes at least one transition point between a straight line and a curve. Accepting input of first and second seed points on both sides of the lesion;
The method according to any preceding claim, wherein the first and second junction points are selected based on the positions of the first and second seed points.   Among the singular points, the first and second ones closest to the first and second seed points are selected as the first and second junction points, respectively. Item 5. The method according to Item 4.   6. The method of claim 5, wherein a dimension from the seed point to the singular point is measured along the boundary to determine the nearest junction point.   The method of claim 6, wherein the seed point is moved toward the boundary to determine a dimension to the singular point along the boundary.   8. If one of the seed points is in the interior space, the seed point is moved relative to the boundary in a direction toward the other one of the seed points. The method described in 1.   If one of the seed points is in the wall, the seed point is moved relative to the boundary such that the angle of rotation of the line between the seed points is limited The method of claim 7.   10. If one of the seed points is in the wall, it is moved relative to the boundary when the distance to the boundary is less than a predetermined threshold. The method described in 1.   For each seed point, the nearest singular point in both directions along the boundary is identified, and the dimension along the boundary to the nearest singular point in both directions satisfies a predetermined criterion, 11. The method according to any one of claims 5 to 10, wherein the closest singularity is selected as the corresponding junction point.   When a line between the first and second junction points intersects the boundary, the first and / or second junction points are mutually aligned along the boundary until the line no longer intersects the boundary. A method according to any preceding claim, characterized in that it is moved towards.   When the line between the first and second junction points protrudes from the lesion, the line is moved toward each other along the boundary until it intersects with the lesion but does not intersect with the boundary. 12. The method according to any one of claims 1 to 11.   Steps a to c are performed on each of a series of consecutive slice images of the scanned image, and the first and second junction points determined in one of the slice images are 12. The method according to any one of claims 4 to 11, wherein the method is used to generate the first and second seed points for one neighboring one.   Steps a to c are performed on each of a series of consecutive slice images of the scanned image, starting from a predetermined starting slice image and in each direction from the starting slice image until the end of the lesion is detected. The method of claim 14, wherein the method is performed.   16. The method of claim 15, wherein an edge of the lesion is detected if the lesion is not detected between the seed points in a current slice image. For each of the series of consecutive slice images, the range of the lesion is detected, and the range of the lesion detected in the current slice image does not overlap the range of the lesion detected in the adjacent slice image. If
The method according to claim 15 or 16, wherein an edge of the lesion is detected. The joint between the lesion and the wall is defined by a joint line extending between the first and second joint points;
A method according to any preceding claim, characterized in that the extent of the lesion is determined to be limited by the junction line.   The method according to claim 18, wherein the extent of the lesion is limited by the connecting line and is determined by dividing the scanned image and performing region expansion.   When the first and second junction points are not connected at the boundary, the range of the lesion is further limited by a boundary connection line that connects the separated boundary sections where the junction points are respectively located. The method according to claim 19.   21. The method of claim 20, wherein the boundary connecting line is defined by determining a change in distance between the separated boundary sections. The extent of the lesion is
Performing a fine separation that separates the image around the lesion into the foreground and background using a threshold based on local contrast;
Identifying a continuous foreground region denoting the lesion, limited by the junction line;
Extending the foreground region to add background points and to exclude foreground points not included in the foreground region to generate a mask;
Obtaining a fuzzy continuity map that defines the continuity of points in the mask relative to seed points in the lesion;
The method according to any one of claims 18 to 21, characterized in that it is determined by determining the extent of a lesion from the fuzzy continuity map.   The lesion range is determined from the fuzzy continuity map by extending a continuous region from a seed point in the fuzzy continuity map and outputting the region having the highest boundary contrast as the lesion range. 23. The method of claim 22, wherein: In the CT scan image of the lung, in the method of identifying the extent of the lesion attached to the wall,
a. Identify the junction between the lesion and the wall;
b. Using a threshold based on local contrast, performing a fine segmentation that separates the image surrounding the lesion into foreground and background;
c. Restricting at the junction, expanding the foreground region showing the lesion,
d. To generate a mask, extend the foreground region to include a background point, excluding foreground points not included in the foreground region,
e. Obtaining a fuzzy continuity map that defines the continuity of points in the mask relative to seed points in the lesion;
f. Determining the extent of the lesion from the fuzzy continuity map. In step f
Extending a continuity region from the seed point in the fuzzy continuity map;
The method according to claim 24, wherein an area having the highest boundary contrast is output as the lesion range.   A method according to any preceding claim, wherein the lesion is a nodule and the internal space is the lung.   The method according to any one of claims 1 to 25, wherein the lesion is a polyp and the internal space is a large intestine.   A computer program configured to perform a method according to any preceding claim.   A computer program device comprising a recording medium for recording the computer program according to claim 28.

Images